Paper deep dive
templar: agentic induction and evolution of standardized radiology reporting templates from large-scale clinical corpora
Xiaotian Hu, Mingxuan Liu, Zhonghan Wang, Xinfeng Zhang, Yiming Huang, Ziang Wang, Kasidit Anmahaepong, Yijin Li, Yifei Chen, Hongjia Yang, Zihan Li, Qiyuan Tian
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/6/2026, 9:52:47 AM
Summary
The paper introduces TEMPLAR, a template-centric agentic framework for inducing and evolving standardized radiology reporting templates from large-scale clinical corpora. TEMPLAR utilizes three agents (Induction, Evolution, Clinical) and two provenance-aware knowledge graphs (anatomical and diagnostic) to overcome limitations of static template induction methods like ASTAR and context-limited single-LLM approaches. TEMPLAR demonstrates superior performance in template coverage, information fidelity, and diagnostic fidelity across four datasets (PETCT, ViMed, MERLIN, BDMAP) compared to ASTAR and various medical and general-purpose LLMs.
Entities (13)
Relation Signals (11)
TEMPLAR → outperforms → ASTAR
confidence 95% · TEMPLAR outperforms ASTAR, three medical LLMs, and six general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity
TEMPLAR → uses → Induction Agent
confidence 95% · The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms
TEMPLAR → uses → Evolution Agent
confidence 95% · the Evolution Agent then assembles these slots into a hierarchical template and revises it
TEMPLAR → uses → Clinical Agent
confidence 95% · and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnostic reasoning.
Clinical Agent → applies → Evolved Template
confidence 90% · and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnostic reasoning.
Evolution Agent → assembles → Hierarchical Template
confidence 90% · the Evolution Agent then assembles these slots into a hierarchical template
Induction Agent → derives → Canonical Clinical Slots
confidence 90% · The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms via dual-view similarity clustering
TEMPLAR → maintains →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Structured radiology reporting mitigates the heterogeneity of free-text reports, yet its benefits depend on high-quality reporting templates. In practice, such templates are conventionally built through labor-intensive expert consensus and therefore vary across institutions and lag behind evolving clinical practice. Large language models (LLMs) enable automated template induction, but existing approaches remain limited: single-LLM induction is constrained by context length, and the corpus-scale method ASTAR produces a static, closed-corpus template without external grounding or downstream adaptation. To address these limitations, we propose TEMPLAR, a TEMPLate-centric Agentic framework for inducing and evolving standardized Radiology reporting templates from large-scale clinical corpora. TEMPLAR treats the template as a persistent central state maintained alongside two provenance-aware knowledge graphs, namely an anatomical graph that constrains template construction and a diagnostic graph that supports finding-to-diagnosis reasoning. Three agents operate on this state. The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms via dual-view similarity clustering; the Evolution Agent then assembles these slots into a hierarchical template and revises it under consistency constraints, external clinical evidence, and downstream structuring feedback; and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnostic reasoning. Across four datasets, TEMPLAR outperforms ASTAR, three medical LLMs, and six general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity, while achieving the highest or tied-highest LLM-rated template quality. Its fidelity advantages over ASTAR persist under cross-dataset transfer, and cumulative ablations support complementary contributions of its key components.
Tags
Links
- Source: https://arxiv.org/abs/2610.05247v1
- Canonical: https://arxiv.org/abs/2610.05247v1
Trouble viewing inline? Open PDF directly →
Full Text
195,890 characters extracted from source content.
Expand or collapse full text
TEMPLAR: AGENTIC INDUCTION AND EVOLUTION OF STANDARDIZED RADIOLOGY REPORTING TEM- PLATES FROM LARGE-SCALE CLINICAL CORPORA Xiaotian Hu 1∗ , Mingxuan Liu 1∗ , Zhonghan Wang 1 , Xinfeng Zhang 2 , Yiming Huang 3 , Ziang Wang 1 , Kasidit Anmahapong 1 , Yijin Li 1 , Yifei Chen 1 , Hongjia Yang 1 , Zihan Li 1 , Qiyuan Tian 1† 1 Tsinghua University 2 University of Hong Kong 3 University of California, San Diego ABSTRACT Structured radiology reporting mitigates the heterogeneity of free-text reports, yet its benefits depend on high-quality reporting templates. In practice, such templates are conventionally built through labor-intensive expert consensus and therefore vary across institutions and lag behind evolving clinical practice. Large language models (LLMs) enable automated template induction, but existing approaches re- main limited: single-LLM induction is constrained by context length, and the corpus-scale method ASTAR produces a static, closed-corpus template without external grounding or downstream adaptation. To address these limitations, we propose TEMPLAR, a TEMPLate-centric Agentic framework for inducing and evolving standardized Radiology reporting templates from large-scale clinical cor- pora. TEMPLAR treats the template as a persistent central state maintained alongside two provenance-aware knowledge graphs, namely an anatomical graph that constrains template construction and a diagnostic graph that supports finding- to-diagnosis reasoning. Three agents operate on this state. The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms via dual-view similarity clustering; the Evolution Agent then assembles these slots into a hierarchical template and revises it under consistency constraints, external clinical evidence, and downstream structuring feedback; and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnos- tic reasoning. Across four datasets (PETCT, ViMed, MERLIN, and BDMAP), TEMPLAR outperforms ASTAR, three medical LLMs, and six general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity, while achieving the highest or tied-highest LLM-rated template quality. Its fidelity advantages over ASTAR persist under cross-dataset transfer, and cumulative ablations sup- port complementary contributions of its key components. 1INTRODUCTION Radiology reports communicate imaging findings for clinical decisions (Ganeshan et al., 2018; Pons et al., 2016) and provide training data for medical artificial intelligence (AI) (Irvin et al., 2019; Pereira et al., 2024). Institutional conventions and individual styles produce inconsistent terminol- ogy, organization, and granularity in free-text reports (Naik et al., 2001; European Society of Radi- ology (ESR), 2018), complicating evidence extraction and normalization for clinical interpretation and AI development (Vosshenrich et al., 2023; Datta et al., 2021; Huesch et al., 2018). Structured reporting first defines a template of clinical fields and value specifications, then populates it from free-text reports (Zhang et al., 2026b;c; Fytas et al., 2024). Although LLMs increasingly au- tomate field population (Chen et al., 2024; Liu et al., 2026), template construction remains largely manual, relying on institutional committees or professional-society consensus (Larson et al., 2013; D’Hondt et al., 2026). Expert involvement preserves clinical relevance, but repeated design, re- view, and adaptation require substantial labor (Busch et al., 2025), limiting deployment across tasks, institutions, and evolving practice. ∗ Equal contribution. † Corresponding author. E-mail: qiyuantian@tsinghua.edu.cn 1 arXiv:2610.05247v1 [cs.MA] 4 Oct 2026 Manually Design DoctorReport Human Labor Time-Consuming Single-LLm Design LLMReport Limited Context Stateless Astar Design ASTARReport Mixed Concepts Static Schema TEMPLAR (Ours) TEMPLARReport Agentic Frame Adaptive Single-LLM Key: lung_fdg_uptake Type: Freetext Value: normal / no abnormal FDG uptake / increased / SUVmax=<NUM> ASTAR Key: nodule_suv_metric Type: Freetext Value: <NUM> Type: Numerical Key: pulmonary_suvmax Value: None Ours Key: nodule_fdg_uptake: Type: Categorical Value: normal / slightly increased / increased Note: normal: SUVmax < 2.5; slightly increased: SUVmax 2.5 –4.0; increased: SUVmax > 4.0 Nodule FDG Update Stare Lung Nodule Normal Increased Slightly Increased Indeterminate Low Metabolism SUVmax Malignancy Suspicion Qualified By Has FDG Takeup Has Degree Consistent With Interpreted As Raises Suspicion For (a) Existing Paradigms Comparison (b) Illustrative Template Comparison Figure 1:Comparison of reporting-template construction paradigms. (a) Labor-intensive manual design, context-limited single-LLM induction, static corpus-scale induction with ASTAR, and adaptive agentic evolution with TEMPLAR. (b) Single-LLM templates may conflate clinical concepts and ASTAR templates retain redundant keys, whereas TEMPLAR couples an evidence- grounded, evolvable template with complementary knowledge graphs. LLMs offer a path toward automated template construction (Adams et al., 2023; Busch et al., 2025; Wo ́znicki et al., 2025), but context limits prevent direct processing of large reporting cor- pora (Modarressi et al., 2025). Truncation and sampling can reduce input length at the cost of incomplete coverage and uneven granularity (Zhang et al., 2026b). ASTAR scales induction by consolidating report-level concepts into canonical slots and a refined anatomical hierarchy (Zhang et al., 2026b). Its closed-corpus templates nevertheless remain static, without systematic external grounding or downstream feedback to detect and address information loss. We propose TEMPLAR, a template-centric agentic framework that induces templates from clinical corpora and maintains them as persistent states. Two provenance-aware knowledge graphs preserve clinical relations: an anatomical graph constrains construction, and a diagnostic graph supports finding-to-diagnosis reasoning (Fig. 1b). The Induction Agent extracts anatomy-constrained Span– Triple atoms, induces slots through dual-view similarity and Leiden clustering (Traag et al., 2019), and consolidates both graphs. The Evolution Agent assembles the hierarchy, revises it under con- sistency constraints and external evidence, and performs one round of feedback-guided granularity calibration. The Clinical Agent applies the template and graphs to structuring, reconstruction, and diagnostic reasoning; the Evaluation Agent assesses template coverage, template quality, informa- tion fidelity, and diagnostic fidelity. Our contributions are threefold: (1) We formulate template construction as agentic induction and evolution, treating the template as a persistent central state coupled with provenance-aware knowl- edge graphs (Fig. 1a). (2) We design an Induction Agent for anatomy-constrained slot induction, an Evolution Agent that refines the template under external evidence and downstream feedback, and a Clinical Agent that applies it to structuring and diagnostic reasoning. (3) On four datasets (PETCT, ViMed, MERLIN, and BDMAP), TEMPLAR achieves higher template coverage, information fi- delity, and diagnostic fidelity than ASTAR and nine LLMs. 2 2RELATED WORK Structured Representations of Medical Reports. Medical report structuring converts free-text narratives into standardized, machine-readable representations, ranging from template-based clin- ical fields to normalized concepts and entity–relation graphs (Kreimeyer et al., 2017; Jain et al., 2021). In radiology, the target structure is often specified through expert-defined reporting systems. For example, PI-RADS, O-RADS, and LI-RADS standardize terminology, categorization, and re- porting for prostate MRI, ovarian and adnexal ultrasound, and liver imaging, respectively (Turkbey et al., 2019; Andreotti et al., 2020; Chernyak et al., 2018). The European Society of Paediatric Ra- diology (ESPR) Fetal Task Force has also developed structured reporting recommendations for fetal MRI (Sofia et al., 2024). With the target fields or templates specified in advance, Transformer- and LLM-based approaches primarily automate their population from free-text reports (Adams et al., 2023; Wo ́znicki et al., 2025). Beyond field-based representations, a complementary line of work explicitly captures relations among clinical concepts through entity extraction, terminology nor- malization, and relation extraction (Datta et al., 2021; Jain et al., 2021). For knowledge-graph construction from unstructured text more broadly, KGGen combines entity–relation extraction with cross-source aggregation and iterative clustering of entities and edges (Mo et al., 2025). Shifting the focus from report content extraction to template induction, ASTAR derives clinical fields and their hierarchical organization from report corpora, reducing reliance on manual template design (Zhang et al., 2026b). However, the resulting template remains a static representation derived from a closed corpus, without external knowledge grounding or adaptation to downstream structuring feedback. TEMPLAR extends this corpus-driven paradigm by incorporating external clinical evidence into template revision and using downstream structuring feedback to calibrate descriptive granularity. Stateful Agentic Systems. Agentic systems combine tools, specialized roles, and persistent state in interactive workflows (Xi et al., 2025; Guo et al., 2024). InnoEval retrieves heterogeneous evidence for research-idea evaluation (Qiao et al., 2026); EvolveR distills interaction trajectories into strate- gies updated through subsequent feedback (Wu et al., 2026); and MedKGent assigns extraction, evidence integration, and graph updates to specialized agents (Zhang et al., 2026a). These systems motivate persistent representations revised through evidence and feedback. TEMPLAR applies this principle to reporting templates: external clinical evidence grounds semantics, structuring feedback calibrates granularity, and structural constraints and provenance support controlled updates. 3METHOD Overview. Given a corpus of free-text radiology reportsD =x i N i=1 , the goal of TEMPLAR (Fig. 2) is to induce and evolve a hierarchical reporting template T ⋆ , where anatomical branches organize clinical slots φ = (k,τ, Ω) specifying a canonical field name k, a value type τ , and a candidate- value set Ω. TEMPLAR obtains T ⋆ throughD A ind −→ Φ keep A evo −→ T ⋆ , where the Induction Agent A ind derives the retained canonical slot library Φ keep from anatomy-constrained Span–Triple atoms, and the Evolution AgentA evo assembles and refines the template using consistency checks, external evidence, and structuring feedback. Two accompanying graphs, G anat and G diag , encode anatomi- cal relations and finding–diagnosis associations, respectively, supporting template construction and downstream clinical applications while preserving source provenance for each canonical relation. Detailed procedures and agent prompts are provided (Appendices B and F). 3.1ANATOMY-GUIDED TEMPLATE INDUCTION Dual-View Similarity Graph. To prevent cross-region concept conflation, A ind segments reports at changes in anatomical focus and groups segments by canonical region. Within each segment, it separately extracts a clinical span u span = (r,ℓ,c,τ,v,ρ) and a relation u kg = (e subj ,p,e obj ). The span records region r, finer location ℓ, attribute c, type τ , value v, and clinical role ρ. The relation specifies subject, predicate, and object. Segment-local alignment forms atoms a = (u span ,u kg ,ξ) with source provenance ξ (Appendices B.1.1 and B.1.2). The full induction procedure is detailed separately (Appendix B.1). 3 Report A Evo Sampled Backbone H 0 Slot Library Initial Assignment Attach / Branch / Drop Coarse Revision Intra-node/Global Revision T 1 Key: Liver Lesion Size Type: Freetext Value: "<num>"; "<num> x <num> cm "<num>cm"; A Evo Fine Revision T f Type: Numerical Value: "<num> cm" A Evo Query Malignant Feature？ Online Evidence TgNote Normal: .. Suspicious: .. A Cli Strc. Reco. A Eva Evaluation Guided Strc. LexiconStrc.Slot Noisy Slot Patterns P s TEMPLATE EVOLUTION Cross-Sourse Evidence Grouding Granularity Calibration Hierarchical Assembly and Revision. A Evo Guided Leiden Cluster λ3 ⋅BM25+λ4 ⋅Embedding Triple Weight Anatomical Triple Diagnostic Triple Span Atom Community Triple-Span Pair SentenceEmbedding The cardiac silhouette appeared normal, with pericardial thickening... A lobulated soft tissue shadow measuring ...in the right middle lobe... ... shows no abnormal density shadows in the liver parenchyma... Liver Lung Heart Report Slot-Level RAG Cluster1Cluster2 λ1 ⋅BM25+λ2 ⋅EmbeddingSpan Weight SlotTriple Slot Induction/Relation Consolidation ANATOMY-GUIDED TEMPLATE INDUCTION Anatomical KG Diagnostic KG RAG Output Figure 2: TEMPLAR overview. Anatomy-guided induction derives clinical slots, which evolve into a template through assembly, revision, external evidence grounding, and structuring feedback. For atom a i , the two views are q text i = concat(r i ,ℓ i ,c i ,v i ) and q graph i = linearize(e subj i ,p i ,e obj i ). Within each region, fused similarities and sparse edge weights are: S ij = X m∈text,graph λ m λ bm25 B m ij + λ emb E m ij , A ij = S ij I S ij ≥ δ ∧ j ∈N k (i)∨ i∈N k (j) . (1) Here, B m is row-normalized, symmetrized BM25 similarity (Robertson & Zaragoza, 2009), E m is cosine similarity between normalized embeddings, the λ terms are fusion weights, andN k (i) con- tains the top-k neighbors of i. Cross-region edges are excluded. Weighted Leiden clustering (Traag et al., 2019) yields anatomically consistent communities of related atoms (Appendix B.1.3). Slot and Relation Consolidation. Each community is partitioned into semantic subgroups, with one slot φ = (k,τ, Ω) created for each subgroup, where k denotes the canonical slot key, τ ∈ categorical,numerical,freetext denotes its value type, and Ω is its candidate-value set. Proposals sharing a normalized key are merged while retaining their evidence. Let A(φ) contain the distinct identifiers of the atoms supporting φ. To suppress incidental concepts, we retain only slots whose distinct-atom support exceeds γ s : Φ keep =φ∈ Φ :|A(φ)| > γ s ,(2) where Φ is the merged slot library. Regional relation consolidation produces G anat while preserving measurement semantics, laterality, and polarity. Separately, paired training findings and impressions yield G diag by normalizing diagnostic targets, including disease and clinical-status concepts, and consolidating finding–target relations with anatomical scope and provenance. Both graphs retain source evidence. Slot induction and relation consolidation are detailed separately (Appendices B.1.4 and B.1.5); representative subgraphs illustrate the resulting relations (Appendix E). Sentence Retrieval Corpus. To provide linguistic exemplars for report reconstruction,A ind indexes normalized, deduplicated training sentences asM sent =(u j ,h j ) M j=1 , whereh j is the normalized embedding of sentence u j . The retrieval procedure is described in Sec. 3.3. 4 3.2TEMPLATE EVOLUTION To improve template consistency and preserve clinical detail, A evo assembles and refines the re- tained slots (Appendix B.2): Φ keep assembly −→ T 1 revision −→ T f grounding −→ T g calibration −→ T ⋆ .(3) The resulting T ⋆ integrates an anatomical hierarchy, evidence-grounded slot annotations, and extrac- tion guidance for preserving clinical detail when structuring individual free-text radiology reports. Hierarchical Assembly and Revision. From representative reports,A evo induces a backbone H 0 . The first assignment round attaches slots or proposes branches; the second reassigns retained slots against the completed hierarchy. Coarse revision through within-node consolidation and cross-path concept harmonization yields T 1 (Appendix B.2.1). Fine revision corrects ambiguous keys, in- compatible types, fragmented vocabularies, and numerical or unit inconsistencies. A deterministic reducer rejects edits violating structural or unit checks. Empty vocabularies are supplemented only when indexed training sentences support coherent values, producing T f . Cross-Source Evidence Grounding. Slots induced from the local report corpus may remain seman- tically underspecified or exhibit dataset-specific biases. Accordingly,A evo queries PubMed, Europe PMC, and Google using each slot’s key, type, vocabulary, and anatomical path in T f . It dedupli- cates documents by identifiers and metadata, ranks them by authority and information richness, and attaches supported evidence as slot-level reference notes. These notes clarify categorical or free-text semantics and provide units, reference intervals, or thresholds, yielding T g (Appendix B.2.2). Behavior-Aware Granularity Calibration. The Clinical AgentA cli (Sec. 3.3) structures sampled training findings with T g . For each slot s, A evo audits its specification φ s , extracted values Y s , aligned source contexts C s , and a template-derived generic-value lexicon V gen . If detailed descrip- tions repeatedly collapse to generic values (e.g., present), the agent derives slot-specific linguistic patterns P s . These patterns augment T g with extraction guidance to form T ⋆ , without changing Ω s . For example, mapping “a 2.1-cm spiculated lesion in the right upper lobe” to present discards its size, morphology, and location. The corresponding P s instructs subsequent extraction to preserve these explicitly reported attributes (Appendix B.2.3). Representative final templates span PETCT and ViMed (Appendices D.1 and D.2) and MERLIN and BDMAP (Appendices D.3 and D.4); their catalog documents the resulting specifications (Appendix D). 3.3CLINICAL STRUCTURING AND REASONING The Clinical AgentA cli uses T ⋆ to structure reports, then supports report reconstruction and diag- nostic reasoning from the resulting structured records. Template-Conditioned Structuring. For a new report x,A cli fills T ⋆ by anatomical region to ob- tain a hierarchical record z. Highly supported, region-relevant triples retrieved from G anat provide priors. Notably,A cli extracts values only from x, leaving unsupported slots unfilled. Slot-Conditioned Reconstruction. For each non-empty slot–value pair (k s ,v s ) ∈ z, A cli con- structs a query and retrieves the closest training sentence: q s = concat(k s ,:,v s ), j s = arg max j cos(Enc(q s ),h j ).(4) After deduplication, the retrieved sentences are ranked by similarity score, and the top-K sentences form the reference set R(z). A cli reconstructs the report ˆx using the structured record z and these reference sentences. The references guide wording and anatomical organization, while all case- specific findings and measurements are drawn from z. Graph-Assisted Diagnostic Reasoning.A cli uses the structured recordzto retrieve relevant finding– target relations from the frozen diagnostic graphG diag . These relations provide training-derived priors for interpreting the findings, inferring a diagnostic conclusion, and identifying the supporting clinical findings. During reasoning,A cli accesses onlyzandG diag , withzsupplying all case-specific evidence. The original report and reference diagnosis are withheld and used only for evaluation. 5 Table 1: Performance comparison on PETCT and ViMed in template coverage, information fidelity, and diagnostic fidelity. Best in bold and runner-up in underlined . Case = case-macro coverage; Key = key-micro coverage; Avg. = average; R-1/R-2/R-L = ROUGE-1/ROUGE-2/ROUGE-L; chrF = character n-gram F-score; chrF++ = character and word n-gram F-score; BERT = BERTScore; PDA = Primary Diagnosis Accuracy; KFP = Key Finding Preservation; CA = Clinical Actionability. Template CoverageInformation FidelityDiagnostic Fidelity MethodCaseKeyAvg.R-1R-2R-LchrF chrF++ BERTAvg.PDAKFPCAAvg. PETCT (n = 49) Medical LLM baselines lingshu-32b0.6678 0.7192 0.6935 0.5920 0.3089 0.3937 0.49160.4687 0.8865 0.5236 0.4939 0.5310 0.6000 0.5416 medgemma-27b0.6744 0.7027 0.6886 0.74950.5163 0.5727 0.62140.60580.90690.66210.5633 0.6051 0.6571 0.6085 clinfusion-32b0.4710 0.4910 0.4810 0.1419 0.0580 0.1089 0.05980.0556 0.7380 0.1937 0.5102 0.5320 0.6204 0.5542 Medical LLM Avg.0.60440.63760.62100.49450.29440.35840.39090.37670.84380.45980.52250.55600.62580.5681 General LLM baselines deepseek-v4-flash0.7632 0.8079 0.7856 0.6736 0.3825 0.4418 0.55550.5327 0.8991 0.5809 0.6245 0.6473 0.7265 0.6661 deepseek-v4-pro0.7565 0.7652 0.7609 0.6863 0.3939 0.4692 0.57480.5524 0.9029 0.5966 0.5959 0.6271 0.6857 0.6362 qwen3.6-flash0.7679 0.7948 0.7814 0.5713 0.2861 0.3607 0.43070.4101 0.8718 0.4885 0.6327 0.6739 0.7265 0.6777 gemini-3.6-flash0.5967 0.5908 0.5938 0.7228 0.53500.58340.58120.5701 0.8954 0.6480 0.6408 0.6876 0.7265 0.6850 gpt-5.40.7683 0.8144 0.7914 0.7117 0.4825 0.5293 0.58310.5688 0.8973 0.6288 0.5878 0.6380 0.6816 0.6358 gemini-3.5-flash-lite 0.6330 0.6450 0.6390 0.5577 0.3449 0.4086 0.41400.3994 0.8808 0.5009 0.6449 0.6654 0.75100.6871 General LLM Avg.0.71430.73630.72530.65390.40420.46550.52320.50560.89120.57390.62110.65660.71630.6647 Template induction baseline ASTAR0.80740.86740.83740.6473 0.3397 0.3863 0.59780.5655 0.8911 0.5713 0.67350.70860.75100.7110 Ours0.87830.90410.89120.81510.60940.58400.77560.75700.92680.74470.72650.75330.80410.7613 VIMED (n = 277) Medical LLM baselines lingshu-32b0.5595 0.5501 0.5548 0.5179 0.2648 0.3881 0.37540.3580 0.8774 0.4636 0.3502 0.4091 0.4188 0.3927 medgemma-27b0.5122 0.5086 0.5104 0.5485 0.2944 0.4038 0.43720.4143 0.8866 0.4975 0.3690 0.4341 0.4390 0.4140 clinfusion-32b0.4557 0.4634 0.4596 0.2440 0.0987 0.1986 0.08490.0785 0.7542 0.2432 0.5588 0.6049 0.6426 0.6021 Medical LLM Avg.0.50910.50740.50830.43680.21930.33020.29920.28360.83940.40140.42600.48270.50010.4696 General LLM baselines deepseek-v4-pro0.5692 0.5693 0.5693 0.5634 0.2798 0.4381 0.42470.3997 0.8856 0.4986 0.3863 0.4600 0.4751 0.4405 qwen3.6-flash0.6522 0.6513 0.6518 0.7101 0.4610 0.5310 0.61820.5877 0.9112 0.6365 0.5935 0.6588 0.6534 0.6352 gemini-3.6-flash0.6384 0.6381 0.6383 0.6044 0.3774 0.4881 0.49460.4722 0.9039 0.5568 0.4498 0.5348 0.5249 0.5032 gpt-5.40.5750 0.5795 0.5773 0.6070 0.3938 0.4993 0.50020.4804 0.9049 0.5643 0.3458 0.4376 0.4231 0.4022 gemini-3.5-flash-lite 0.7283 0.7335 0.7309 0.77200.59210.64620.73780.71830.93000.73270.6325 0.6928 0.70320.6762 deepseek-v4-flash0.7293 0.7325 0.7309 0.5648 0.2829 0.4326 0.44650.4273 0.8883 0.5071 0.5025 0.5495 0.5841 0.5454 General LLM Avg.0.64870.65070.64970.63700.39780.50590.53700.51430.90400.58270.48510.55560.56060.5338 Template induction baseline ASTAR0.84890.85180.85040.6339 0.3147 0.4180 0.56660.5332 0.8955 0.5603 0.64690.70330.6931 0.6811 Ours0.88150.87580.87870.90660.78610.75390.87420.86800.96090.85830.65780.73180.71480.7015 3.4MULTI-DIMENSIONAL EVALUATION The Evaluation AgentA eval assesses template quality and downstream performance (Appendix B.3). Template quality is quantified by averaging the scores for design quality (13 dimensions) and clin- ical utility (6 dimensions), computed separately from LLM-as-judge (Zheng et al., 2023) and expert ratings (Appendix B.3.2). Template coverage measures how well template slots represent clinical concepts in held-out reports through LLM-based semantic matching, with coverage aggregated at the case-macro and key-micro levels (Appendix B.3.1). Information fidelity quantifies the agree- ment between reconstructed reports ˆx and the original held-out findings using ROUGE-1/2/L (Lin, 2004), chrF/chrF++ (Popovi ́ c, 2015; 2017), and BERTScore (Zhang* et al., 2020), all computed directly against the references (Appendix B.3.3). Diagnostic fidelity assesses the preservation of diagnostically relevant information through LLM ratings of primary diagnosis accuracy (PDA), key finding preservation (KFP), and clinical actionability (CA), using the held-out diagnoses and find- ings as references. All three metrics are reported on a [0, 1] scale (Appendix B.3.4). 4EXPERIMENTS AND RESULTS 4.1EXPERIMENTAL SETUP Datasets. We evaluate TEMPLAR on four English-language radiology report datasets cover- ing two distinct imaging settings, namely whole-body 18 F-FDG PET/CT and abdominal CT. The PET/CT datasets comprise PETCT (Xue et al., 2026) with 490 reports (441/49 train/test) and ViMed (Nguyen et al., 2025) with 2,761 reports (2,484/277). The abdominal CT datasets comprise MERLIN (Blankemeier et al., 2026) with 4,080 reports (3,672/408) and BDMAP (Bassi et al., 2025) 6 Table 2: Performance comparison on MERLIN and BDMAP in template coverage, information fidelity, and diagnostic fidelity. Best in bold and runner-up in underlined . Template CoverageInformation FidelityDiagnostic Fidelity MethodCaseKeyAvg.R-1R-2R-LchrF chrF++ BERTAvg.PDAKFPCAAvg. MERLIN (n = 408) Medical LLM baselines lingshu-32b0.5214 0.5347 0.5281 0.5677 0.3047 0.3765 0.49650.4701 0.8521 0.5113 0.86620.89750.86370.8758 medgemma-27b0.3230 0.3801 0.3516 0.6072 0.3891 0.3860 0.47050.4538 0.8766 0.5305 0.7137 0.8057 0.7456 0.7550 clinfusion-32b0.3293 0.3435 0.3364 0.2769 0.1159 0.1589 0.20840.1947 0.7833 0.2897 0.5853 0.6650 0.6333 0.6279 Medical LLM Avg.0.39120.41940.40530.48390.26990.30710.39180.37290.83730.44380.72170.78940.74750.7529 General LLM baselines deepseek-v4-flash0.4111 0.4556 0.4334 0.4970 0.2116 0.2791 0.35660.3384 0.8553 0.4230 0.6956 0.7632 0.7225 0.7271 deepseek-v4-pro0.5563 0.5854 0.5709 0.4962 0.2143 0.2855 0.44920.4179 0.8466 0.4516 0.7348 0.8016 0.7637 0.7667 qwen3.6-flash0.5505 0.6082 0.5794 0.5745 0.3059 0.3596 0.44840.4273 0.8657 0.4969 0.7480 0.8157 0.7657 0.7765 gemini-3.6-flash0.67240.70020.68630.6750 0.4555 0.4379 0.53440.5152 0.8819 0.5833 0.7583 0.8413 0.7868 0.7955 gpt-5.40.5738 0.6334 0.6036 0.69790.50330.47990.56630.54980.88770.61420.7426 0.8152 0.7681 0.7753 gemini-3.5-flash-lite 0.5411 0.6103 0.5757 0.5751 0.3497 0.3849 0.43340.4175 0.8711 0.5053 0.6892 0.7695 0.7162 0.7250 General LLM Avg.0.55090.59890.57490.58600.34000.37110.46470.44440.86810.51240.72810.80110.75380.7610 Template induction baseline ASTAR0.6546 0.6918 0.6732 0.5866 0.2967 0.3346 0.48470.4594 0.8641 0.5044 0.8431 0.8825 0.8520 0.8592 Ours0.68330.71680.70010.79190.63420.53870.68720.67600.89820.70440.89310.94090.89660.9102 BDMAP (n = 200) Medical LLM baselines lingshu-32b0.3364 0.3316 0.3340 0.3909 0.2214 0.3091 0.19730.1998 0.8920 0.3684 0.2340 0.2310 0.2390 0.2347 medgemma-27b0.6456 0.6278 0.6367 0.6477 0.4416 0.5244 0.41990.4307 0.9178 0.5637 0.6380 0.6861 0.6510 0.6584 clinfusion-32b0.7014 0.6742 0.6878 0.7041 0.4668 0.4965 0.51880.5213 0.9226 0.6050 0.8400 0.8778 0.8530 0.8569 Medical LLM Avg.0.56110.54450.55280.58090.37660.44330.37870.38390.91080.51240.57070.59830.58100.5833 General LLM baselines deepseek-v4-flash0.6730 0.6543 0.6637 0.7557 0.52040.57690.51900.5298 0.93070.63880.8440 0.8796 0.8660 0.8632 deepseek-v4-pro0.6857 0.6714 0.6786 0.7060 0.5128 0.5677 0.52810.5334 0.9268 0.6291 0.7980 0.8229 0.8200 0.8136 qwen3.6-flash0.3756 0.3664 0.3710 0.2308 0.1351 0.1884 0.11010.1126 0.8712 0.2747 0.1870 0.1870 0.2090 0.1943 gemini-3.6-flash0.6374 0.6091 0.6233 0.6665 0.4643 0.5488 0.44420.4535 0.9215 0.5831 0.6070 0.6540 0.6260 0.6290 gpt-5.40.4338 0.4368 0.4353 0.4412 0.3090 0.3452 0.24710.2511 0.9039 0.4163 0.2770 0.2707 0.2860 0.2779 gemini-3.5-flash-lite 0.6742 0.6535 0.6639 0.6823 0.4718 0.5463 0.48320.4879 0.9242 0.5993 0.6300 0.6693 0.6570 0.6521 General LLM Avg.0.58000.56520.57260.58040.40220.46220.38860.39470.91310.52350.55720.58060.57730.5717 Template induction baseline ASTAR0.8641 0.88850.87630.7333 0.4609 0.4861 0.55820.55520.9129 0.6178 0.86900.91940.88100.8898 Ours0.90110.90830.90470.92020.85080.83990.87310.87400.97050.88810.98100.98700.91000.9593 with 2,000 reports (1,800/200). Template induction, knowledge-graph construction, and method- specific adaptation use only the training split, with the test split held out for evaluation. We use findings for template construction, structuring, and reconstruction, reserving training impressions for diagnostic-graph construction and test impressions for diagnostic evaluation (Appendix C.1). Baselines. We compare TEMPLAR with ASTAR (Zhang et al., 2026b) and nine LLMs. (1) general-purpose LLMs: DeepSeek-V4-Flash, DeepSeek-V4-Pro (DeepSeek-AI et al., 2026), Qwen3.6-Flash (Qwen Team, 2026), Gemini-3.6-Flash (Google DeepMind, 2026a), GPT-5.4 (OpenAI, 2026), and Gemini-3.5-Flash-Lite (Google DeepMind, 2026b); and (2) medical LLMs: Lingshu-32B (LASA Team et al., 2025), MedGemma-27B (Sel- lergren et al., 2025), and ClinFusion-32B (Yuan et al., 2026). To accommodate context limits, each LLM baseline processes the training reports in batches and merges valid local templates into a single template (Appendix C.2). Implementation settings are specified separately (Appendix A). Additional model screening and construction failures (Appendices C.6 and C.6.1), parsing and re- covery (Appendices C.6.2 and C.6.3), and repetition (Appendix C.6.4) are documented. 4.2COMPARISON RESULTS Template quality. TEMPLAR consistently ranked among the best-performing methods in both expert and LLM assessments of overall template quality (Fig. 3). Under expert evaluation, it ranked first on PETCT, ViMed, and MERLIN (4.250 each), while tying with ASTAR and GPT- 5.4 on BDMAP (3.750), below Gemini-3.5-Flash-Lite (4.000). Under LLM evaluation, it tied with MedGemma-27B and ClinFusion-32B on PETCT (2.560) and ranked first on ViMed (2.435), MER- LIN (2.315), and BDMAP (2.520). Its margins over the strongest baseline on the latter three datasets were 0.115, 0.440, and 0.605, respectively. Template coverage. TEMPLAR achieved broader coverage of clinical concepts in held-out reports across all four datasets (Tables 1 and 2). It ranked first at both the case and key levels, reaching 7 1.52.53.54.5 ClinFusion-32B MedGemma-27B Lingshu-32B DeepSeek-V4-Flash DeepSeek-V4-Pro Qwen3.6-Flash Gemini-3.6-Flash Gemini-3.5-Flash-Lite GPT-5.4 ASTAR Ours (a) PETCT 2.500 2.500 3.250 2.500 2.750 3.000 3.000 4.000 3.750 4.000 4.250 1.52.53.54.5 (b) ViMed 2.500 2.750 2.750 2.500 2.750 3.000 3.000 4.000 3.750 4.000 4.250 1.52.53.54.5 (c) MERLIN 2.000 2.250 2.500 3.500 2.500 3.000 3.000 4.000 3.750 3.500 4.250 1.52.53.54.5 (d) BDMAP 3.250 3.000 2.500 3.000 3.500 2.500 2.750 4.000 3.750 3.750 3.750 1.01.52.02.5 ClinFusion-32B MedGemma-27B Lingshu-32B DeepSeek-V4-Flash DeepSeek-V4-Pro Qwen3.6-Flash Gemini-3.6-Flash Gemini-3.5-Flash-Lite GPT-5.4 ASTAR Ours 2.560 2.560 2.200 2.280 2.200 2.435 2.045 2.440 2.320 2.435 2.560 1.01.52.02.5 2.165 1.925 1.925 2.320 1.960 2.040 1.885 2.240 2.000 2.200 2.435 1.01.52.02.5 1.355 1.480 1.680 1.605 1.760 1.875 1.760 1.760 1.560 1.550 2.315 1.01.52.02.5 1.160 1.240 1.160 1.800 1.800 1.240 1.320 1.435 1.320 1.915 2.520 Expert-rated LLM-rated Overall template-quality score Figure 3: Template quality across four datasets. Expert ratings (top) and LLM ratings (bottom) are shown for (a) PETCT, (b) ViMed, (c) MERLIN, and (d) BDMAP. a macro-average coverage of 0.8437, representing a gain of 3.44 percentage points over ASTAR (0.8093), the strongest baseline on this aggregate (Table 5). Notably, TEMPLAR showed its largest coverage advantage on the PETCT dataset, scoring 0.8912 versus 0.8374 for ASTAR. Information fidelity. TEMPLAR reconstructed reports more faithfully across all four datasets, with gains spanning lexical, character-level, and semantic measures (Tables 1 and 2). It ranked first on all six reconstruction metrics and achieved a macro-average information fidelity of 0.7988, exceeding Gemini-3.6-Flash, the strongest baseline (0.5928), by 34.8% (Table 5). Diagnostic fidelity. TEMPLAR achieved the highest diagnostic fidelity across all four datasets (Tables 1 and 2). It ranked first in primary diagnosis accuracy, key finding preservation, and clinical actionability, reaching a macro-average score of 0.8331 compared with 0.7853 for ASTAR, the strongest baseline on this aggregate (Table 5). Notably, TEMPLAR showed its largest diagnostic- fidelity advantage on BDMAP, scoring 0.9593 versus 0.8898 for ASTAR (Table 2). Supplementary results include cross-dataset aggregates and detailed template-quality comparisons (Appendices C, C.3, and C.5). Cov_case Cov_key ROUGE-1 ROUGE-2 ROUGE-L chrF chrF++ BERTScore PDA KFP CA Metric 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Score template coverage Information FidelityDiagnostic Fidelity Open circles: ASTAR | Filled circles: Ours PETCT → ViMed ASTARPETCT → ViMed OursMERLIN → BDMAP ASTARMERLIN → BDMAP Ours 0.4414 0.4434 0.5503 0.8492 0.4367 0.4376 0.5594 0.8410 0.5812 0.8635 0.5160 0.7397 0.3535 0.7849 0.2519 0.5389 0.4043 0.7654 0.3330 0.5595 0.4078 0.8213 0.4857 0.7375 0.4020 0.8205 0.4571 0.7217 0.9034 0.9602 0.8790 0.9251 0.5250 0.6780 0.4884 0.6166 0.5605 0.7250 0.5463 0.6888 0.5870 0.7140 0.5674 0.6773 Figure 4: Cross-dataset generalization. Template coverage, information fidelity, and diagnostic fidelity under PETCT→ViMed and MERLIN→BDMAP transfer, without target-domain template re-induction. Open and filled circles denote ASTAR and TEMPLAR, respectively. 8 4.3CROSS-DATASET GENERALIZATION We compared the cross-dataset generalization of TEMPLAR and ASTAR in two transfer settings: PETCT→ViMed and MERLIN→BDMAP (Fig. 4). Specifically, templates induced on each source dataset were applied directly to held-out reports from the corresponding target dataset, without target-domain template re-induction. For PETCT→ViMed, TEMPLAR achieved case/key cov- erage of 0.8492/0.8410, exceeding ASTAR’s 0.5503/0.5594. It also outperformed ASTAR in re- construction and diagnostic fidelity, with ROUGE-1 of 0.7397 versus 0.5160 and PDA of 0.6166 versus 0.4884. For MERLIN→BDMAP, TEMPLAR achieved comparable case/key coverage to ASTAR (0.4414/0.4367 vs. 0.4434/0.4376), while attaining higher ROUGE-2 (0.7849 vs. 0.3535), BERTScore (0.9602 vs. 0.9034), and PDA (0.6780 vs. 0.5250), highlighting benefits beyond broader coverage alone. Together, these results support the cross-dataset reuse of TEMPLAR tem- plates, with higher reconstruction and diagnostic fidelity observed both alongside broader target- domain coverage (PETCT→ViMed) and at coverage comparable to ASTAR (MERLIN→BDMAP). 4.4ABLATION STUDIES M1M2M3M4M5M6 0.8074 0.8783 0.8674 0.9041 0.8489 0.8815 0.8518 0.8758 0.5503 0.8492 0.5594 0.8410 Case Key Case Key Case Key PETCT ViMed OOD Mean gain (p) PETCT +5.4 ViMed +2.8 OOD +29.0 (a) 0.6473 0.6275 0.6466 0.6700 0.8151 0.3397 0.3341 0.3491 0.3941 0.6094 0.3863 0.3781 0.3953 0.4219 0.5840 0.6339 0.6780 0.7641 0.8570 0.9066 0.3147 0.4034 0.5394 0.6835 0.7861 0.4180 0.4905 0.5938 0.6980 0.7539 0.5160 0.4995 0.6353 0.6466 0.7397 0.2519 0.2283 0.4373 0.4365 0.5389 0.3330 0.3270 0.4411 0.4545 0.5595 R-1 R-2 R-L R-1 R-2 R-L R-1 R-2 R-L PETCT ViMed OOD Mean gain (p) PETCT +21.2 ViMed +36.0 OOD +24.6 (b) 0.5978 0.6226 0.6256 0.6499 0.7756 0.5655 0.5890 0.5920 0.6193 0.7570 0.8911 0.8858 0.8858 0.9002 0.9268 0.5666 0.6412 0.7321 0.8016 0.8742 0.5332 0.6131 0.7124 0.7931 0.8680 0.8955 0.9086 0.9246 0.9499 0.9609 0.4857 0.4665 0.6633 0.6657 0.7375 0.4571 0.4363 0.6448 0.6475 0.7217 0.8790 0.8787 0.9023 0.9072 0.9251 chrF chrF++ BERT chrF chrF++ BERT chrF chrF++ BERT PETCT ViMed OOD Mean gain (p) PETCT +13.5 ViMed +23.6 OOD +18.8 (c) 0.6735 0.6490 0.6612 0.6694 0.7265 0.7086 0.6755 0.6873 0.7084 0.7533 0.7510 0.7388 0.7551 0.7633 0.8041 0.6469 0.6556 0.6130 0.6491 0.6578 0.7033 0.7191 0.6820 0.7226 0.7318 0.6931 0.7191 0.6881 0.7097 0.7148 0.4884 0.4303 0.5343 0.5805 0.6166 0.5463 0.4870 0.5948 0.6538 0.6888 0.5674 0.5141 0.6029 0.6513 0.6773 PDA KFP CA PDA KFP CA PDA KFP CA PETCT ViMed OOD Mean gain (p) PETCT +5.0 ViMed +2.0 OOD +12.7 (d) Figure 5: Ablation study results. Template coverage (a), information fidelity (b,c), and diagnostic fidelity (d) on PETCT, ViMed, and PETCT→ViMed transfer (OOD). Starting from the baseline (M1), M2–M6 successively add the Dual-View Similarity Graph, Anatomical KG Structuring, Slot- Conditioned RAG, Template Evolution, and Diagnostic KG Reasoning. We assessed component contributions of TEMPLAR on PETCT, ViMed, and PETCT→ViMed transfer (Fig. 5). Starting from the baseline (M1, i.e., ASTAR), we successively added the Dual- View Similarity Graph (M2), Anatomical KG Structuring (M3), Slot-Conditioned RAG (M4), Tem- plate Evolution (M5), and Diagnostic KG Reasoning (M6), with each configuration retaining all preceding components. Coverage is reported for M1–M2 and reconstruction for M1–M5, as subse- quent additions leave the respective metrics unchanged. Diagnostic fidelity includes M1–M3, M5, and M6; M4 is omitted because its retrieval step affects reconstruction alone. The induction stage expanded concept coverage, most markedly under PETCT→ViMed transfer, where M2 achieved case/key coverage of 0.8492/0.8410 versus 0.5503/0.5594 for M1. Yet M2 had lower ROUGE-1 and PDA than M1 on both PETCT and PETCT→ViMed, indicating that broader coverage alone did not ensure higher reconstruction or diagnostic fidelity. Subsequent stages sup- ported more faithful reconstruction. On PETCT, ROUGE-1 reached 0.6466 with anatomical guid- ance (M3) and 0.6700 with retrieval (M4), compared with 0.6275 for M2. Template Evolution (M5) yielded the largest incremental ROUGE-1 gain in this setting, reaching 0.8151. M5 also attained ROUGE-1/BERTScore of 0.9066/0.9609 on ViMed and 0.7397/0.9251 under PETCT→ViMed transfer, with the latter exceeding M4’s 0.6466/0.9072. Finally, Diagnostic KG Reasoning (M6) achieved higher PDA than M5 on PETCT (0.7265 vs. 0.6694), ViMed (0.6578 vs. 0.6491), and PETCT→ViMed (0.6166 vs. 0.5805). These findings support complementary roles for concept coverage, information preservation, and diagnostic reasoning in carrying clinical evidence from free-text reports into useful structured representations. Additional ablations examine template revi- sion and evidence notes in terms of template quality and clinical utility (Appendix C.4). 9 5CONCLUSION We presented TEMPLAR, a template-centric agentic framework for inducing and evolving stan- dardized radiology reporting templates from clinical corpora. It integrates anatomy-guided in- duction, external evidence grounding, and structuring feedback with anatomical and diagnostic knowledge graphs for report structuring and reasoning. Across four datasets, TEMPLAR achieved broader template coverage and higher reconstruction and diagnostic fidelity than ASTAR and nine LLM baselines. Its fidelity advantages over ASTAR persisted under cross-dataset transfer, while cumulative ablations supported complementary contributions from its components. By enabling scalable template construction, TEMPLAR could support clearer communication of diagnostic ev- idence for clinical decision-making and facilitate downstream clinical AI model development. AI USE STATEMENT We used generative AI tools for language polishing, manuscript formatting checks, and reference sketches that informed the design of some figures. The authors made all final decisions regarding the manuscript, figures, and scientific content and take full responsibility for the accuracy and integrity of this paper, including its experimental results, numerical values, and bibliographic references. ETHICS STATEMENT This work investigates automated template construction and report processing using existing ra- diology report datasets, as described in Appendix C.1. Differences in data provenance, imaging modality, reporting practice, and language may introduce biases and limit generalization. In par- ticular, translated reports and automatically generated reports may contain artifacts that affect both template construction and evaluation. The framework is intended for research purposes. Generated templates, clinical reference notes, reconstructed reports, and diagnostic outputs may omit relevant information or contain unsupported interpretations. Automated metrics and expert assessments in this study do not establish clinical safety or readiness for deployment. Any clinical application would require independent validation in the intended setting and oversight by qualified clinicians. Generated outputs should not be used as a substitute for professional clinical judgment. REPRODUCIBILITY STATEMENT The appendices document the experimental setup and methodological details to support repro- ducibility. Appendix C.1 describes the datasets, report selection, and training/test split sizes. Ap- pendix A specifies the models, computing environment, and implementation hyperparameters. Ap- pendix B details the framework, and Appendix C.2 describes batched template construction for the LLM baselines. Evaluation procedures are described in Appendix B.3, and the provided prompt specifications appear in Appendix F. Structured-output failures and fallback recovery are discussed in Appendices C.6.2 and C.6.3. Representative template specifications are presented in Appendix D. REFERENCES Lisa C. Adams, Daniel Truhn, Felix Busch, Avan Kader, Stefan M. Niehues, Marcus R. Makowski, and Keno K. Bressem. Leveraging gpt-4 for post hoc transformation of free-text radiology reports into structured reporting: A multilingual feasibility study. Radiology, 307(4):e230725, 2023. doi: 10.1148/radiol.230725. Rochelle F. Andreotti, Dirk Timmerman, Lisa M. Strachowski, Wouter Froyman, Beryl R. Benacer- raf, Gary L. Bennett, Tom Bourne, Deborah L. Brown, Beverly G. Coleman, Mary C. Frates, Steven R. Goldstein, Ulrike M. Hamper, Mindy M. Horrow, Marta Hernanz-Schulman, Car- oline Reinhold, Stephen L. Rose, Brian P. Whitcomb, Wendy L. Wolfman, and Paula Glanc. O-RADS US risk stratification and management system: A consensus guideline from the ACR ovarian-adnexal reporting and data system committee. Radiology, 294(1):168–185, 2020. doi: 10.1148/radiol.2019191150. 10 Pedro R.A.S. Bassi, Mehmet Can Yavuz, Ibrahim Ethem Hamamci, Sezgin Er, Xiaoxi Chen, Wenx- uan Li, Bjoern Menze, Sergio Decherchi, Andrea Cavalli, Kang Wang, Yang Yang, Alan Yuille, and Zongwei Zhou. Radgpt: Constructing 3d image-text tumor datasets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 23720–23730, October 2025. L. Blankemeier, A. Kumar, J. P. Cohen, et al. Merlin: a computed tomography vision–language foundation model and dataset. Nature, 652:1318–1328, 2026. doi: 10.1038/s41586-026-10181-8. Felix Busch, Lena Hoffmann, Daniel Pinto dos Santos, et al. Large language models for structured reporting in radiology: past, present, and future. European Radiology, 35:2589–2602, 2025. doi: 10.1007/s00330-024-11107-6. Yuxuan Chen, Haoyan Yang, Hengkai Pan, Fardeen Siddiqui, Antonio Verdone, Qingyang Zhang, Sumit Chopra, Chen Zhao, and Yiqiu Shen. Burextract-llama: An llm for clinical concept extrac- tion in breast ultrasound reports. In Proceedings of the 1st International Workshop on Multimedia Computing for Health and Medicine, MCHM’24, p. 53–58, New York, NY, USA, 2024. Asso- ciation for Computing Machinery. ISBN 9798400711954. doi: 10.1145/3688868.3689200. URL https://doi.org/10.1145/3688868.3689200. Victoria Chernyak, Kathryn J. Fowler, Aya Kamaya, Ania Z. Kielar, Khaled M. Elsayes, Mustafa R. Bashir, Yuko Kono, Richard K. G. Do, Donald G. Mitchell, Amit G. Singal, An Tang, and Claude B. Sirlin. Liver imaging reporting and data system (LI-RADS) version 2018: Imag- ing of hepatocellular carcinoma in at-risk patients. Radiology, 289(3):816–830, 2018. doi: 10.1148/radiol.2018181494. S. Datta, J. Godfrey-Stovall, and K. Roberts. Radlex normalization in radiology reports. In AMIA Annual Symposium Proceedings, volume 2020, p. 338–347. American Medical Informatics As- sociation, 2021. DeepSeek-AI, Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chengyu Hou, Chenhao Xu, Chenze Shao, Chong Ruan, Conner Sun, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Donghao Li, Dongjie Ji, Erhang Li, Fang Wei, Fangyun Lin, Fangzhou Yuan, Feiyu Xia, Fucong Dai, Guangbo Hao, Guanting Chen, Guoai Cao, Guolai Meng, Guowei Li, Han Yu, Han Zhang, Hanwei Xu, Hao Li, Haofen Liang, Haoling Zhang, Haoming Luo, Haoran Wei, Haotian Yuan, Haowei Zhang, Haowen Luo, Haoyu Chen, Haozhe Ji, Hengqing Zhang, Honghui Ding, Hongxuan Tang, Huanqi Cao, Huazuo Gao, Hui Qu, Hui Zeng, J Yang, JQ Zhu, Jia Luo, Jia Song, Jia Yu, Jialiang Huang, Jialu Cai, Jian Liang, Jiangting Zhou, Jiasheng Ye, Jiashi Li, Jiaxin Xu, Jiewen Hu, Jieyu Yang, Jin Chen, Jin Yan, Jingchang Chen, Jingli Zhou, Jingting Xiang, Jingyang Yuan, Jingyuan Cheng, Jingzi Zhou, Jinhua Zhu, Jiping Yu, Joseph Sun, Jun Ran, Junguang Jiang, Junjie Qiu, Junlong Li, Junmin Zheng, Junxiao Song, Kai Dong, Kaige Gao, Kang Guan, Kexing Zhou, Kezhao Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Wang, Leyi Xia, Li Zhang, Liang Zhao, Lihua Guo, Lingxiao Luo, Linwang Ma, Linyan Zhu, Litong Wang, Liyu Cai, Liyue Zhang, Longhao Chen, MS Di, MY Xu, Max Mei, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Mingxu Zhou, Minmin Han, Ning Wang, Panpan Huang, Panpan Wang, Peixin Cong, Peiyi Wang, Peng Zhang, Qiancheng Wang, Qihao Zhu, Qingyang Li, Qinyu Chen, Qiushi Du, Qiwei Jiang, Rui Tian, Ruifan Xu, Ruijie Lu, Ruiling Xu, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, Runqian Chen, Runqiu Yin, Runxin Xu, Ruomeng Shen, Ruoyu Zhang, Ruyi Chen, SH Liu, Shanghao Lu, Shangmian Sun, Shangyan Zhou, Shanhuang Chen, Shaofei Cai, Shaoheng Nie, Shaoqing Wu, Shaoyuan Chen, Shengding Hu, Shengyu Liu, Shiqiang Hu, Shirong Ma, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, Shuying Yu, Songyang Zhou, Tao Ni, Tao Yun, Tian Jin, Tian Pei, Tian Ye, Tianle Lin, Tianran Ji, Tianyi Cui, Tianyuan Yue, Tingting Yu, Tun Wang, W Zhang, WL Xiao, Wangding Zeng, Wei An, Weilin Zhao, Wen Liu, Wenfeng Liang, Wenjie Pang, Wenjing Luo, Wenjing Yao, Wenjun Gao, Wenkai Yang, Wenlve Huang, Wenqing Hou, Wentao Zhang, Wenting Ma, Xi Gao, Xiang He, Xiangwen Wang, Xianzu Wang, Xiao Bi, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaokang Zhang, Xiaotao Nie, and Xiaowen Sun. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv e-prints, art. arXiv:2606.19348, April 2026. doi: 10.48550/arXiv.2606.19348. 11 Aurélie D’Hondt, Marie Cassart, Elka Miller, Willemijn M. Klein, Susan Shelmerdine, Michael Aertsen, Eléonore Blondiaux, Marianne Alison, Ajay Taranath, Christian Abel, Monica Re- bollo Polo, Marta Gomez-Chiari, Teresa Victoria, Padma Rao, David Perry, Lizbet Perez-Marrero, Sudha A. Anupindi, Manisha Jana, Andrea Levai, Giana Izzo, Megan B. Marine, Rick R. van Rijn, Stacy Goergen, and Owen J. Arthurs. Standardised structured fetal and neonatal post- mortem magnetic resonance imaging reporting: recommendations from the european society of paediatric radiology postmortem task force. Pediatric Radiology, 56(7):1480–1488, 2026. doi: 10.1007/s00247-026-06634-x. European Society of Radiology (ESR). Esr paper on structured reporting in radiology. Insights into Imaging, 9(1):1–7, 2018. doi: 10.1007/s13244-017-0588-8. Panagiotis Fytas, Anna Breger, Ian Selby, Simon Baker, Shahab Shahipasand, and Anna Korhonen. Can rule-based insights enhance LLMs for radiology report classification? introducing the Rad- Prompt methodology. In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, p. 212–235, 2024. doi: 10.18653/v1/2024.bionlp-1.17. Dhakshinamoorthy Ganeshan, Phuong T. Duong, Linda Probyn, Leon Lenchik, Tara A. McArthur, Michele Retrouvey, Elham H. Ghobadi, Sarah L. Desouches, David Pastel, and Isaac R. Francis. Structured reporting in radiology. Academic Radiology, 25(1):66–73, 2018. doi: 10.1016/j.acra. 2017.08.005. Google DeepMind. Gemini 3.6 Flash model card. Google DeepMind, jul 2026a. URL https: //deepmind.google/models/model-cards/gemini-3-6-flash/. Google DeepMind. Gemini 3.5 flash-lite model card. Google DeepMind, jul 2026b. URL https: //deepmind.google/models/model-cards/gemini-3-5-flash-lite/. Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: a survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial In- telligence, IJCAI ’24, 2024. ISBN 978-1-956792-04-1. doi: 10.24963/ijcai.2024/890. URL https://doi.org/10.24963/ijcai.2024/890. Marco D. Huesch, R. Cherian, S. Labib, and R. Mahraj. Evaluating report text variation and infor- mativeness: Natural language processing of ct chest imaging for pulmonary embolism. Journal of the American College of Radiology, 15(3):554–562, 2018. doi: 10.1016/j.jacr.2017.12.017. Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Christopher Chute, Henrik Marklund, Behzad Haghgoo, Robyn L. Ball, Katie S. Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Cur- tis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y. Ng. CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison. In AAAI Confer- ence on Artificial Intelligence, p. 590–597, 2019. doi: 10.1609/AAAI.V33I01.3301590. URL https://mlanthology.org/aaai/2019/irvin2019aaai-chexpert/. Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew P. Lungren, Andrew Y. Ng, Curtis Langlotz, and Pranav Ra- jpurkar. Radgraph: Extracting clinical entities and relations from radiology reports. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. URL https://openreview.net/forum?id=pMWtc5NKd7V. Kory Kreimeyer, Matthew Foster, Abhishek Pandey, Nina Arya, Gwendolyn Halford, Sandra F. Jones, Richard Forshee, Mark Walderhaug, and Taxiarchis Botsis. Natural language processing systems for capturing and standardizing unstructured clinical information: A systematic review. Journal of Biomedical Informatics, 73:14–29, 2017. doi: 10.1016/j.jbi.2017.07.012. David B. Larson, Alex J. Towbin, Rebecca M. Pryor, and Lane F. Donnelly. Improving consistency in radiology reporting through the use of department-wide standardized structured reporting. Ra- diology, 267(1):240–250, 2013. doi: 10.1148/radiol.12121502. 12 LASA Team, Weiwen Xu, Hou Pong Chan, Long Li, Mahani Aljunied, Ruifeng Yuan, Jianyu Wang, Chenghao Xiao, Guizhen Chen, Chaoqun Liu, Zhaodonghui Li, Yu Sun, Junao Shen, Chaojun Wang, Jie Tan, Deli Zhao, Tingyang Xu, Hao Zhang, and Yu Rong. Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning. arXiv e-prints, art. arXiv:2506.07044, June 2025. doi: 10.48550/arXiv.2506.07044. Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, p. 74–81, Barcelona, Spain, July 2004. Association for Computational Linguis- tics. URL https://aclanthology.org/W04-1013/. Mingxuan Liu, Yijin Li, Juncheng Zhu, Hongjia Yang, Yiming Huang, Haoxiang Li, Yifei Chen, Xuguang Bai, Yi Liao, Haibo Qu, and Qiyuan Tian. Fetalextract-llm: Structured information extraction from free-text fetal mri reports based on privacy-ensuring open-weights large language models. In Perinatal, Preterm and Paediatric Image Analysis, p. 119–129, Cham, 2026. Springer Nature Switzerland. ISBN 978-3-032-05997-0. Belinda Mo, Kyssen Yu, Joshua Kazdan, Proud Mpala, Lisa Yu, Charilaos I. Kanatsoulis, and Sanmi Koyejo. KGGen: Extracting knowledge graphs from plain text with language models. In Ad- vances in Neural Information Processing Systems, volume 38, p. 33962–33985, 2025. doi: 10.52202/085713-1010. Ali Modarressi, Hanieh Deilamsalehy, Franck Dernoncourt, Trung Bui, Ryan A. Rossi, Seunghyun Yoon, and Hinrich Schuetze. NoLiMa: Long-context evaluation beyond literal matching. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu (eds.), Proceedings of the 42nd International Con- ference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, p. 44554–44570. PMLR, 13–19 Jul 2025. URL https://proceedings.mlr.press/ v267/modarressi25a.html. S. S. Naik, A. Hanbidge, and S. R. Wilson. Radiology reports: examining radiologist and clinician preferences regarding style and content. American Journal of Roentgenology, 176(3):591–598, 2001. doi: 10.2214/ajr.176.3.1760591. Tien Nguyen, Dac Nguyen, Duc Nguyen The Minh, Trung Thanh Nguyen, Truong Thao Nguyen, Hieu Pham, Johan Barthelemy, Tran Minh Quan, Quoc Viet Hung Nguyen, Thanh Tam Nguyen, Mai Son, Chau Anh, Thanh Nguyen, and Phi Le Nguyen. Toward a vision-language founda- tion model for medical data: Multimodal dataset and benchmarks for vietnamese PET/CT re- port generation. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-3907. OpenAI. GPT-5.4. OpenAI, mar 2026. URL https://developers.openai.com/api/ docs/models/gpt-5.4. S. C. Pereira, A. M. Mendonça, A. Campilho, P. Sousa, and C. Teixeira Lopes. Automated image label extraction from radiology reports: A review. Artificial Intelligence in Medicine, 149:102814, 2024. doi: 10.1016/j.artmed.2024.102814. Ewoud Pons, Loes M. M. Braun, M. G. Myriam Hunink, and Jan A. Kors.Natural lan- guage processing in radiology: a systematic review. Radiology, 279(2):329–343, 2016. doi: 10.1148/radiol.16142770. Maja Popovi ́ c. chrF: character n-gram F-score for automatic MT evaluation. In Ond ˇ rej Bojar, Rajan Chatterjee, Christian Federmann, Barry Haddow, Chris Hokamp, Matthias Huck, Varvara Logacheva, and Pavel Pecina (eds.), Proceedings of the Tenth Workshop on Statistical Machine Translation, p. 392–395, Lisbon, Portugal, September 2015. Association for Computational Lin- guistics. doi: 10.18653/v1/W15-3049. URL https://aclanthology.org/W15-3049/. Maja Popovi ́ c. chrF++: words helping character n-grams. In Ond ˇ rej Bojar, Christian Buck, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Ji- meno Yepes, Philipp Koehn, and Julia Kreutzer (eds.), Proceedings of the Second Conference on Machine Translation, p. 612–618, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/W17-4770. URL https://aclanthology. org/W17-4770/. 13 Shuofei Qiao, Yunxiang Wei, Xuehai Wang, Bin Wu, Boyang Xue, Ningyu Zhang, Hossein A. Rahmani, Wang Yanshan, Qiang Zhang, Keyan Ding, Jeff Z. Pan, Huajun Chen, and Emine Yil- maz. InnoEval: On research idea evaluation as a knowledge-grounded, multi-perspective rea- soning problem. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=8AhAziAWvZ. Qwen Team.Qwen3.6-Flash.Alibaba Cloud Model Studio, 2026.URL https://w. alibabacloud.com/help/en/model-studio/qwen3-6-flash. Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3(4):333–389, April 2009. ISSN 1554-0669. doi: 10.1561/1500000019. URL https://doi.org/10.1561/1500000019. Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, Justin Chen, Fereshteh Mahvar, Liron Yatziv, Tiffany Chen, Bram Sterling, Stefanie Anna Baby, Susanna Maria Baby, Jeremy Lai, Samuel Schmidgall, Lu Yang, Kejia Chen, Per Bjornsson, Shashir Reddy, Ryan Brush, Ken- neth Philbrick, Mercy Asiedu, Ines Mezerreg, Howard Hu, Howard Yang, Richa Tiwari, Sunny Jansen, Preeti Singh, Yun Liu, Shekoofeh Azizi, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Riviere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean-bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Elena Buchatskaya, Jean-Baptiste Alayrac, Dmitry Lepikhin, Vlad Feinberg, Sebastian Borgeaud, Alek Andreev, Cassidy Hardin, Robert Dadashi, Léonard Hussenot, Armand Joulin, Olivier Bachem, Yossi Matias, Katherine Chou, Avinatan Hassidim, Kavi Goel, Clement Farabet, Joelle Barral, Tris Warkentin, Jonathon Shlens, David Fleet, Vic- tor Cotruta, Omar Sanseviero, Gus Martins, Phoebe Kirk, Anand Rao, Shravya Shetty, David F. Steiner, Can Kirmizibayrak, Rory Pilgrim, Daniel Golden, and Lin Yang. MedGemma Technical Report. arXiv e-prints, art. arXiv:2507.05201, July 2025. doi: 10.48550/arXiv.2507.05201. C. Sofia, M. Aertsen, C. Garel, and M. Cassart. Standardised and structured reporting in fetal magnetic resonance imaging: recommendations from the fetal task force of the european so- ciety of paediatric radiology. Pediatric Radiology, 54(10):1566–1578, 2024. doi: 10.1007/ s00247-024-06010-7. Vincent A. Traag, Ludo Waltman, and Nees Jan van Eck. From louvain to leiden: guaranteeing well- connected communities. Scientific Reports, 9:5233, 2019. doi: 10.1038/s41598-019-41695-z. Baris Turkbey, Andrew B. Rosenkrantz, Masoom A. Haider, Anwar R. Padhani, Geert Villeirs, Katarzyna J. Macura, Clare M. Tempany, Peter L. Choyke, Francois Cornud, Daniel J. Margolis, Harriet C. Thoeny, Sadhna Verma, Jelle Barentsz, and Jeffrey C. Weinreb. Prostate imaging reporting and data system version 2.1: 2019 update of prostate imaging reporting and data system version 2. European Urology, 76(3):340–351, 2019. doi: 10.1016/j.eururo.2019.02.033. J. Vosshenrich, I. Nesic, D. T. Boll, and T. Heye. Investigating the impact of structured reporting on the linguistic standardization of radiology reports through natural language processing over a 10- year period. European Radiology, 33(11):7496–7506, 2023. doi: 10.1007/s00330-023-10050-2. P. Wo ́znicki, C. Laqua, I. Fiku, et al. Automatic structuring of radiology reports with on-premise open-source large language models. European Radiology, 35:2018–2029, 2025. doi: 10.1007/ s00330-024-11074-y. Rong Wu, Xiaoman Wang, Jianbiao Mei, Pinlong Cai, Daocheng Fu, Cheng Yang, Licheng Wen, Xuemeng Yang, Yufan Shen, Yuxin Wang, and Botian Shi. From interactions to princi- ples: Experience-driven self-distillation for evolving LLM agents. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id= BamDfCP5R3. Zhiheng Xi, Wenxiang Chen, Xin Guo, et al. The rise and potential of large language model based agents: a survey. Science China Information Sciences, 68:121101, 2025. doi: 10.1007/ s11432-024-4222-0. 14 L. Xue, G. Feng, W. Zhang, et al. PETWB-REP: A multi-cancer whole-body FDG PET/CT dataset with corresponding radiology reports. Scientific Data, 13:675, 2026. doi: 10.1038/ s41597-026-07058-w. Hangjie Yuan, Yichen Qian, Zhiwei Tang, Xianzhe Xu, Lirong Wu, Sicheng Yang, Jinwang Wang, Pengju Wang, Zhitao Zeng, Yizeng Han, Yan Xing, Shengxuan Luo, Tao Feng, Qing Xie, Weigen Yao, Yi Yang, Zuozhu Liu, Jiasheng Tang, Shaocheng Wang, Jitao Wang, Jiahong Dong, Weihua Chen, Feng Xu, and Fan Wang. ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding. arXiv e-prints, art. arXiv:2607.24743, July 2026. doi: 10. 48550/arXiv.2607.24743. Duzhen Zhang, Zixiao Wang, Zhong-Zhi Li, Yahan Yu, Shuncheng Jia, Jiahua Dong, Haotian Xu, Xing Wu, Yingying Zhang, Tielin Zhang, Jie Yang, Xiuying Chen, and Le Song. MedKGent: a large language model agent framework for constructing temporally evolving medical knowledge graph. npj Digital Medicine, 2026a. doi: 10.1038/s41746-026-03058-7. Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=SkeHuCVFDr. Xinfeng Zhang, Mingxuan Liu, Yifei Chen, Juncheng Zhu, Kasidit Anmahapong, Yiming Huang, Yuan Zhang, Hongjia Yang, Yi Liao, Gang Ning, Haibo Qu, and Qiyuan Tian. ASTAR: Auto- mated induction of STAndardized radiology Reporting templates from large-scale clinical free- text corpora. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026, 2026b. Yuan Zhang, Xinfeng Zhang, Xiaoming Qi, Xinyu Wu, Feng Chen, Guanyu Yang, and Huazhu Fu. Content generation models in computational pathology: A comprehensive survey on methods, applications, and challenges. IEEE Reviews in Biomedical Engineering, 19:374–395, 2026c. doi: 10.1109/RBME.2025.3619086. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URL https: //openreview.net/forum?id=uccHPGDlao. 15 AIMPLEMENTATION DETAILS AND HYPERPARAMETER SETTINGS Models and computing environment. Unless otherwise specified, the Induction Agent and Evolu- tion Agent use Qwen3.6-Flash for anatomy segmentation, semantic-atom induction, slot construc- tion, and template refinement. The Clinical Agent uses Gemini-3.5-Flash for template-conditioned report structuring and slot-conditioned report reconstruction. The Evaluation Agent uses DeepSeek- V4-Flash for clinical-concept extraction and semantic concept-to-slot matching. GPT-5.5 serves as the LLM judge for the automated assessment of Template Design Quality and Clinical Utility. Diag- nostic outputs were generated using Gemini-3.5-Flash and evaluated using deepseek-v4-flash. All model calls use a temperature of 0. Experiments were conducted on a server equipped with NVIDIA A800 GPUs (80GB). The GPUs were used primarily for BERTScore evaluation. Embeddings and multi-view graph construction. We use text-embedding-3-small to en- code clinical spans, relation triples, and training sentences. The embeddings are 1,536-dimensional, and corpus embeddings are computed in batches of 128. Within each text or relation view, row- normalized BM25 similarity and cosine embedding similarity are combined with equal weights of 0.5. BM25 uses k 1 = 1.5, b = 0.75, and ε = 0.25. The text and relation views are then fused with weights of 0.4 and 0.6. For each anatomical region, we construct a symmetric k-nearest- neighbor graph using k = 20 and an edge threshold of δ = 0.25. Leiden partitioning uses the Reichardt–Bornholdt configuration model, a resolution of γ = 1.5, and random seed 0, and is run until convergence. Regions containing fewer than 30 atoms are retained as single communities. Slot induction and external evidence grounding. A slot is retained when it is supported by more than γ s = 3 distinct atoms, corresponding to at least four supporting atoms. Candidate categorical values are ranked by support frequency, and at most 20 values are retained for each slot. The anatomical backbone is induced from 100 sampled training reports. For evidence grounding, the Evolution Agent queries PubMed, Europe PMC, and Google. It retrieves up to K r = 8 records from each source, deduplicates the combined pool, and provides the six highest-ranked records for evidence-note synthesis. Retrieved documents are truncated to 6,000 characters, the evidence text supplied from each selected record is truncated to 600 characters, and each slot-level reference note is limited to 200 characters. Sentence retrieval and metrics. The sentence corpus is constructed from normalized and dedu- plicated training sentences. Sentences shorter than 15 characters or longer than 400 characters are discarded. For each non-empty slot-value pair, the Clinical Agent retrieves its nearest training sen- tence by cosine similarity. Retrieved sentences are deduplicated and ranked, and the top 80 sentences are supplied as linguistic and organizational references for reconstruction. ROUGE-1, ROUGE-2, and ROUGE-L are reported using F-measure with stemming. chrF and chrF++ are computed at the sentence level, with word order 2 for chrF++. BERTScore uses roberta-large, layer 17, batch size 16, and a maximum sequence length of 512, and reports the F1 score. BMETHOD DETAILS B.1TEMPLATE INDUCTION B.1.1ANATOMICAL-ORDER PRIOR AND ORDERED SEGMENTATION Medical reports often exhibit a weak anatomical-order prior, whereby findings from the same anatomical region tend to appear contiguously. We use this prior to prevent lexically similar findings from distinct anatomical regions from being incorrectly linked during subsequent global processing. For each report x i , the Induction AgentA ind partitions the text into contiguous segments according to their anatomical regions (Fig. 36): G i =A seg ind (x i ) =g ij M i j=1 , g ij = ( e b ij ,x ij ),(5) where x ij denotes a contiguous excerpt from the original report and e b ij is its anatomical region identified by the agent. Each segment retains its original text and position in the report. Because reports are segmented independently, the same anatomical region may receive different labels across reports. For example, chest, thorax, and thoracic region may all refer to the same 16 region. We therefore standardize variations in capitalization, spacing, formatting, and terminology by mapping equivalent labels to a shared English snake-case name, such as chest. After label standardization, semantic atoms associated with the same anatomical region are pooled across reports. Similarity computation, neighborhood construction, and community detection are then performed independently within each region. Semantic atoms from different anatomical regions are not matched, and their cross-region affinities are fixed to zero. B.1.2SEGMENT-LOCAL SPAN–RELATION ALIGNMENT Clinical spans encode fine-grained attributes and values, whereas relation triples capture the rela- tional context among clinical entities. Since the two views are extracted independently (Fig. 37), A ind aligns them to construct unified semantic atoms for subsequent dual-view graph construction. For a segment g, let U g =u span s n g s=1 , R g = n u kg t o m g t=1 (6) denote its ordered span collection and relation set, respectively. For each span, the Induction Agent selects the most semantically relevant relation from the same segment according to the span fields and the relation’s subject, predicate, and object (Fig. 38): t s =A align ind (u span s ,R g ), a s = u span s ,u kg t s ,ξ s .(7) The resulting atoms preserve the number and order of the input spans, with ξ s retaining source provenance. A relation may be assigned to multiple spans, while the alignment encourages coverage of distinct segment-level relations before reuse whenever they are semantically compatible. B.1.3ANATOMY-CONSTRAINED MULTI-VIEW GRAPH AND LEIDEN PARTITION ASTAR applies K-means to globally cluster dense embeddings of clinical spans by minimizing the squared distance between each span and its assigned cluster centroid: J KM = X i ∥h i −μ c i ∥ 2 2 ,(8) whereh i is the embedding of span i andμ c i is the centroid of its assigned cluster. This objec- tive represents each clinical concept cluster by a single prototype and measures cluster coherence through proximity to that prototype. It is therefore better suited to concept groups that form compact distributions around representative centroids in a shared vector space, but can be poorly matched to clinical concepts with heterogeneous local densities. In contrast, our atom similarity is not derived from a single vector distance. It integrates sparse lexical similarity and dense semantic similarity from both the text and relation views, producing heterogeneous pairwise affinities that cannot nec- essarily be reduced to Euclidean distances from a single centroid without losing information. To preserve the multi-view information encoded by these pairwise affinities, we directly represent the fused affinity S ij as a weighted graph. Let b i ∈B denote the normalized anatomical region of atom a i , and let V b =i| b i = b(9) denote the atoms assigned to region b. Graph construction is restricted to atoms within the same region, such that A ij = 0 whenever b i ̸= b j . For each i ∈ V b , we define its regional k-nearest- neighbor set according to the fused text–relation affinity in Eq. (1): N (b) k (i) = TopK j∈V b \i S ij .(10) The directed neighborhoods are then converted into a sparse undirected graph: A (b) ij = S ij I [S ij ≥ δ] I h j ∈ N (b) k (i) ∨ i∈ N (b) k (j) i , i,j ∈V b .(11) The neighborhood-union rule preserves a strong one-sided nearest-neighbor relation when local concept densities are asymmetric, while the threshold δ removes weak edges that could otherwise connect unrelated concepts. We use k = 20 and δ = 0.25. 17 For the resulting regional graph, let d (b) i = X j∈V b A (b) ij ,2w b = X i∈V b d (b) i (12) denote the weighted degree of node i and twice the total undirected edge weight, respectively. For w b > 0, we define the degree-corrected excess affinity between two atoms as ∆ (b) ij = A (b) ij − γ d (b) i d (b) j 2w b ,(13) where the second term is their expected connection strength under a degree-preserving null model. Leiden community detection (Traag et al., 2019) then approximately maximizes the Reichardt– Bornholdt objective: b c (b) ≈ arg max c (b) 1 2w b X i,j∈V b ∆ (b) ij I h c (b) i = c (b) j i .(14) This objective identifies atom communities whose observed internal affinities exceed those expected from their weighted degrees. We set the resolution parameter to γ = 1.5. Anatomical regions containing fewer than n min = 30 atoms bypass graph partitioning and are retained as single com- munities because the available evidence is insufficient for a stable finer partition. For regions with at least n min atoms, the Leiden objective applies when w b > 0, with isolated atoms retained as singleton communities. If w b = 0, every atom is isolated; the stated singleton convention therefore defines the partition without evaluating the objective. Letting b C b denote the resulting partition of region b, the complete partition is b C = [ b∈B b C b .(15) Each community retains its member atom identifiers and source provenance for subsequent slot induction. Unlike K-means, which groups atoms according to their proximity to a shared centroid and requires the selection of a global cluster count C, our construction uses anatomical labels to restrict the admissible comparison scope and multi-view evidence to determine local connections. Leiden then groups atoms according to whether their observed affinities exceed the expectations of a degree-preserving null model, allowing the number and structure of communities to adapt to each anatomical region. This combination reduces the risk that anatomically distinct or semantically heterogeneous clinical concepts are absorbed into the same global centroid-based cluster. B.1.4COMMUNITY-TO-SLOT TRANSITION Let C = a i |C| i=1 denote a Leiden community, where a i is its i-th semantic atom and |C| is the number of atoms in the community. Although atoms within the same community are similar in anatomical location, textual semantics, and relational structure, they may encode distinct clinical attributes. The slot-induction action therefore identifies clinically meaningful semantic distinctions within C and divides it into Q C semantically coherent subgroups, where q ∈ 1,...,Q C indexes the q-th subgroup. One canonical slot is induced from each subgroup (Fig. 39): φ C,q = (k C,q ,τ C,q , Ω C,q ),(16) where k C,q is the canonical English slot key, τ C,q ∈ categorical,numerical,freetext is its value type, and Ω C,q is the candidate-value set supported by the subgroup evidence. For example, one community in the induced graph contains 65 adrenal-gland atoms, including 44 atoms describing SUVmax, 13 describing size, and two describing diameter. Taking the SUVmax subgroup as an example, its corresponding slot is φ C,2 = (adrenal_suvmax,numerical,⟨NUM⟩).(17) This example illustrates that atoms from the same anatomical region may belong to a common Leiden community while still requiring separate slots for anatomically related but clinically distinct measurements, such as structural dimensions and metabolic activity. A single slot is induced by default when a community concentrates on one clinical attribute; multiple slots are produced only when the community contains a clinically meaningful semantic separation. 18 After community-level induction, slot proposals with the same canonical key are merged globally. Candidate values are first deduplicated and then ranked by the number of supporting proposals, with the most frequently supported values retained. When proposals disagree on value type, the majority type is used as the initial merged type. For keys supported by multiple proposals, an LLM may further harmonize the canonical name, description, value type, and candidate-value set without changing the underlying atom-to-slot evidence assignments (Fig. 41). Let A(φ) denote the set of distinct atoms associated with slot φ. Slot support is defined as|A(φ)|, and only slots satisfying the support threshold are retained: Φ keep =φ∈ Φ||A(φ)| > γ s .(18) We set γ s = 3, requiring each retained slot to be supported by at least four distinct atoms. Only the filtered slot library Φ keep is passed to the hierarchical construction stage described in Ap- pendix B.2.1. B.1.5PROVENANCE-PRESERVING RELATION CONSOLIDATION Relation normalization is performed in parallel with slot induction. Within each community C, the subject, predicate, and object of each relation triple are normalized, and semantically equivalent relations are mapped to a canonical triple (Fig. 40) ̄g = ( ̄ h, ̄p, ̄ t). Across communities, whitespace, capitalization, and delimiters are first standardized. Triples with identical normalized forms are then merged deterministically according to their complete signatures: σ(g) = (norm(h), norm(p), norm(t)).(19) Semantic comparison is subsequently restricted to triples associated with the same normalized anatomical region. Subject to predefined clinical constraints, the LLM merges relations that use different entity or predicate expressions but convey equivalent clinical semantics (Fig. 42). These constraints preserve clinically meaningful distinctions. Measurement relations retain their mea- surement targets and quantitative semantics and are not collapsed into generic attribute relations. Relations involving left- and right-sided anatomical structures remain separate to preserve laterality. Relations representing positive findings, negative findings, and normal states are also kept distinct to prevent changes in clinical polarity during consolidation. After consolidation, each semantic-equivalence class is represented by a canonical triple ̄g = ( ̄ h, ̄p, ̄ t). The canonical triple retains its normalized anatomical region, source atom identifiers, and source community identifiers, thereby preserving its correspondence to the original evidence. During clinical structuring, these canonical triples provide region-specific relational priors over re- current anatomy–finding–attribute patterns. B.2TEMPLATE EVOLUTION B.2.1DETAILED HIERARCHICAL TEMPLATE CONSTRUCTION Anatomical Backbone Induction. To construct the initial hierarchy, the Evolution Agent A evo samples a fixed number of training reports and provides their description fields to the LLM. The LLM is instructed to generate a nested JSON tree with desc as its unique root (Fig. 46). Internal nodes organize the reports by major anatomical regions and organ-level structures, whereas leaf nodes remain empty for subsequent slot attachment. The resulting anatomical backbone H 0 defines the candidate paths used during the first round of slot assignment. First-Round Slot Assignment. For each retained slot φ j = (k j ,τ j , Ω j ), the first-round assignment jointly considers its normalized key, semantic description, value type, and candidate vocabulary (Fig. 47). The LLM returns a (1) j = d (1) j ,t (1) j ,p (1) j ,b (1) j ,(20) where d (1) j ∈ attach,propose_new,drop specifies whether the slot is attached to an ex- isting path, requires a new branch, or is discarded. The tier t (1) j ∈ core,extended,drop 19 indicates whether the slot belongs to the core template, the extended template, or neither. The candi- date path is denoted by p (1) j , and b (1) j optionally specifies a proposed branch through its parent path and branch name. A proposed branch is accepted only if its parent path exists and its branch name is valid. Duplicate branches are ignored. Accepted branches are incorporated into H 0 to produce the updated backbone H 1 . Because all first-round assignments are performed independently against the initial path set, a slot cannot use a branch proposed by another slot in the same round. We therefore introduce a second global reassignment stage. Global Path Reassignment. All non-dropped slots are re-evaluated against the complete path set of H 1 (Fig. 48). This stage does not create additional branches. Instead, it selects or corrects each slot’s path within the completed hierarchy. Slots assigned to the core or extended tier are attached to their selected nodes, whereas dropped slots and assignments without valid paths are excluded. The resulting hierarchy and slot assignments form the preliminary template T pre . Semantic Reorganization and Concept Harmonization. After path assignment,A evo refines the template in two steps. First, slots within the same anatomical node are jointly examined to identify semantically redundant fields and strongly overlapping candidate vocabularies. Their candidate values are consolidated and their value types are reconciled while the original anatomical path is preserved (Fig. 49). The published intra-node prompt uses substantial candidate-value overlap as its merging criterion and majority voting to reconcile types. Value overlap alone does not establish semantic identity. Second, slots with the same normalized key across different anatomical paths are aligned in naming, candidate vocabulary, and value type while retaining their anatomical locations. These operations preserve the assigned anatomical paths and produce the coarsely revised template T 1 . Fine Revision. Starting from T 1 , fine revision addresses ambiguous slot keys, incompatible value types, fragmented vocabularies, and numerical or unit inconsistencies. Candidate edits that fail the structural or unit checks of the deterministic reducer are rejected. Empty vocabularies are sup- plemented only when indexed training sentences support coherent values. The accepted revisions produce T f , which is passed to evidence grounding. B.2.2CROSS-SOURCE EVIDENCE GROUNDING For each slot s ∈ T f in the refined template, the Evolution Agent A evo constructs a type-aware retrieval query from the slot specification and its anatomical context (Fig. 50): q s = QueryGen (k s ,τ s , Ω s ,π s ),(21) where k s , τ s , Ω s , and π s denote the slot key, value type, current vocabulary, and anatomical path, respectively. A bilingual slot label is also provided when available. The retrieval objective depends on the slot type. For numerical slots, A evo searches for accepted units, reference intervals, and abnormality thresholds. For categorical and free-text slots, it searches for terminology and descriptions that distinguish normal, benign, and pathological states within the anatomical context. If query generation returns an empty result, the system constructs a fallback query by combining the slot key with the phrase normal range clinical. Let B = PubMed, Europe PMC, Google denote the external evidence sources. The initial evi- dence pool is defined as e R s = [ b∈B Retrieve b (q s ,K r ), K r = 8,(22) where up to eight records are retrieved from each source. Available abstracts and full texts are retained, with each document truncated to at most 6,000 characters. Records returned by multiple sources are deduplicated using DOI, PMID, PMCID, normalized URL, and normalized title. When two records refer to the same document, their identifiers are merged, and the version with richer textual evidence is retained. In this subsection, q s denotes the external-evidence search query for slot s, rather than the slot-value query used for report reconstruction. The deduplicated documents are ranked first by source authority and then by information richness. The source priority is PubMed> Europe PMC> Google. For a documentd, its information-richness 20 score is defined as ρ(d) = 2I fulltext (d) + I abstract (d) + min |b d | 500 , 5 ,(23) where b d denotes the textual evidence contained in the document. This ranking prioritizes authorita- tive sources and, within the same source, favors records containing full text, an abstract, or a richer evidence body. The resulting ranked records constitute the candidate evidence set R s for slot s. Notably, ρ(d) denotes document information richness, rather than the clinical-role field ρ of a span. The top six records are provided to A evo , with the evidence body of each record truncated to 600 characters. The agent then condenses the retrieval query and candidate evidence into a slot-level reference note (Fig. 51). Each note is written in English and limited to 200 characters. For numerical slots, the note summarizes evidence-informed units, reference intervals, and abnormality thresholds, prioritizing values explicitly reported in the retrieved snippets. For categorical and free-text slots, it summarizes the main clinical distinctions among normal, benign, and pathological states. The generated reference note is attached to the slot as auxiliary metadata without replacing its key, value type, or vocabulary. The grounded template is therefore constructed as T g = T f ⊕s.note← n s s∈T f ,(24) where n s denotes the reference note associated with slot s. If no usable evidence is retrieved or note generation fails, the original slot is retained without a note. The system also records the retrieval query, evidence provenance, generation rationale, and execution status for auditing. B.2.3BEHAVIOR-AWARE GRANULARITY CALIBRATION. Although compact slot vocabularies facilitate standardized extraction, they may collapse detailed clinical descriptions into generic states. After the Clinical AgentA cli applies T g to sampled training reports, the Evolution AgentA evo audits the resulting structured records Z =z i . For each slot s, it examines the slot specification (k s ,τ s , Ω s ), the extracted values Y s , their aligned source contexts C s , and a template-induced generic-value lexicon V gen , which contains generic categorical values such as present and absent. A slot is flagged when fine-grained source descriptions are repeatedly mapped to such generic values. For each flagged slot, A evo summarizes the fine-grained information lost during structuring into a compact set of slot-specific linguistic patterns P s . For example, if descriptions containing explicit measurements are repeatedly mapped to present, P s guides subsequent structuring to preserve the exact measurement and unit, as well as explicitly stated anatomical location, lesion morphology, and other clinically meaningful attributes. These patterns provide slot-level guidance for source- grounded structuring without modifying the candidate-value set Ω s . B.3MULTI-DIMENSIONAL EVALUATION DETAILS Template Design Quality and Clinical Utility are assessed separately by LLM-based and expert evaluators, whereas Template Coverage and Diagnostic Fidelity use LLM judgments anchored to held-out reports and diagnostic references. Information Fidelity further provides a deterministic reference-based assessment without LLM involvement. Together, these complementary evaluations provide multiple perspectives on template quality and downstream performance, reducing reliance on any single LLM-as-judge assessment. B.3.1TEMPLATE COVERAGE For the i-th held-out report, let C i denote the set of extracted clinical concept keys (Fig. 55), and let M (c,T ) ∈ 0, 1 indicate whether concept c receives lexical or semantic coverage in template T (Fig. 56). Given an evaluation set of N eval reports, case-macro coverage is defined as Cov case = 1 N eval N eval X i=1 P c∈C i M (c,T ) |C i | .(25) This metric first computes the coverage ratio for each report and then averages the resulting ratios, thereby assigning equal weight to each held-out case. 21 Concept-micro coverage is defined as Cov key = P N eval i=1 P c∈C i M (c,T ) P N eval i=1 |C i | .(26) Unlike case-macro coverage, Cov key assigns equal weight to each concept occurrence across the complete evaluation set. B.3.2TEMPLATE QUALITY AND CLINICAL UTILITY EVALUATION A eval evaluates each template from two complementary perspectives: Template Design Quality and Clinical Utility. Both the LLM-as-judge and domain experts assess the same 19 dimensions, com- prising 13 dimensions of template design quality and six dimensions of clinical utility. The complete evaluation criteria specify both aspects (Tables 3 and 4). The identifiers TD1–TD13 and DQ1–DQ13 refer to the same 13 template-design dimensions in the same order. The same 19 evaluation dimensions and scoring scale were retained across datasets, while the evalu- ator prompts and expert forms were instantiated with dataset-specific modality and task descriptions: for abdominal CT, PET/CT-specific wording was replaced with abdominal-CT terminology, includ- ing the imaging modality (TD1), oncology-related abnormality examples such as pancreatic tumors, vascular invasion, and tumor staging (TD2), and examination ranges such as single-organ, upper- abdominal, and whole-abdominal CT (TD10); although the source reports were in English, TD5 evaluated both the English slot names and their required Chinese glosses (k zh ). Table 3: Dimensions of Template Design Quality. The 13 dimensions assess the structural design, clinical safety, adaptability, and overall usability of a radiology reporting template. The criteria shown below correspond to the PET/CT instantiation. CategoryIDEvaluation criterion Coverage and completeness TD1Coverage of the essential structures and measurements required in PET/CT reports. Coverage and completeness TD2Explicit support for common abnormalities, including primary tumors, lymph-node metastases, and distant metastases involving the bone, liver, or lung. Structure and granularity TD3Hierarchical organization consistent with radiological reasoning, proceeding from findings to organ systems and key anatomical structures. Structure and granularity TD4Appropriate field granularity that preserves clinically relevant information without making the template unnecessarily fragmented or burdensome to complete. Terminology and clarity TD5Clear field names and Chinese labels that conform to commonly used clinical terminology. Terminology and clarity TD6Natural and usable permissible values, such as Normal, Abnormal, Uncertain, and NotAssessed. Safety and error prevention TD7Ability of the template to reduce the risk of omitting clinically important findings. Safety and error prevention TD8Protection against incorrectly recording an unassessed structure as normal, including natural support for the NotAssessed state. Safety and error prevention TD9Appropriate representation of diagnostic uncertainty, including expressions such as consider, possible, and cannot be excluded. Adaptability and extensibility TD10Adaptability to different examination ranges, including regional, torso, and whole-body PET/CT examinations. Deployment feasibility TD11Feasibility of using the template in routine clinical reporting. Sign-off safety TD12Whether the template is sufficiently safe to support clinical report sign-off from a template-design perspective. Overall assessment TD13Overall clinical usability of the template after considering all preceding dimensions. 22 Table 4: Dimensions of Clinical Utility. The 6 dimensions assess the expected value of the template in routine radiological interpretation and reporting workflows. The criteria shown below correspond to the PET/CT instantiation. CategoryIDEvaluation criterion Omission reduction CU1The template reduces omissions of clinically important findings. Reporting consistency CU2The template improves consistency across PET/CT reports. Workflow alignment CU3The ordering of template fields is consistent with the clinician’s image-reading workflow. Unassessed-state support CU4The template enables unassessed structures or findings to be represented naturally using the NotAssessed state. Communication efficiency CU5The template improves communication efficiency between radiologists and referring clinicians. Overall satisfaction CU6Overall evaluator satisfaction with the template. Score aggregation. For each evaluator type e ∈ LLM, Expert, we first average the correspond- ing dimension-level ratings to obtain the Template Design Quality score S (e) design and the Clinical Utility score S (e) utility : S (e) design = 1 13 13 X i=1 s (e) DQ,i , S (e) utility = 1 6 6 X j=1 s (e) CU,j .(27) Here s (e) DQ,i and s (e) CU,j denote the ratings of design-quality dimension i (DQ1–DQ13) and clinical- utility dimension j (CU1–CU6), respectively. The overall template-quality score for evaluator type e is then computed as S (e) template = 1 2 S (e) design + S (e) utility .(28) B.3.3INFORMATION FIDELITY Information fidelity measures the preservation of patient-specific information throughout the com- plete text→ structure→ text cycle. Specifically, given an original report x, the Clinical Agent first converts it into a hierarchical structured record z using the final template T ⋆ (Fig. 52) and then gen- erates a reconstructed report ˆx from z (Fig. 53). The reconstructed report ˆx is compared with the original report x using ROUGE-1, ROUGE-2, ROUGE-L, chrF, chrF++, and BERTScore. These metrics characterize reconstruction fidelity from lexical, character-level, and contextual-semantic perspectives. The structured record z remains the sole source of patient-specific facts during recon- struction. Retrieved training sentences provide only linguistic and organizational priors and must not introduce patient-specific findings absent from z. B.3.4DIAGNOSTIC FIDELITY Diagnostic fidelity is evaluated under an information-closure setting. During diagnosis generation (Fig. 54), the Clinical Agent may access only the structured representation z of the current case and the frozen diagnostic graph G diag constructed exclusively from the training set. The reference report and diagnosis annotation are withheld during reasoning and used only for final evaluation. The Evaluation Agent acts as an LLM judge, scoring the generated diagnostic output against the withheld reference report and diagnosis annotation. The judging prompt specifies diagnostic-output evaluation (Fig. 57). We assess three complementary criteria: PDA is rated on a 0–5 scale: 0, absent or wholly incorrect; 1, nonspecific but compatible with the reference; 2, partially correct information but an incorrect primary diagnosis; 3, correct primary diagnosis with a clinically important qualifier missing or incorrect; 4, correct diagnosis with only minor discrepancies; 5, clinically equivalent to the reference, including essential qualifiers and di- agnostic uncertainty. Important qualifiers include anatomical site, laterality, and severity or stage when specified in the reference. 23 KFP is the proportion of reference key findings correctly preserved: KFP = M/K, where K is the number of distinct diagnostically relevant reference findings and M is the number correctly retained. The reference findings and their essential qualifiers are fixed before inspecting any method’s output and shared across methods. Semantically equivalent expressions count as matches; omissions, con- tradictions, and missing essential qualifiers do not. Unsupported additions receive no credit. Cases with K = 0 are excluded from KFP aggregation, with the effective sample size reported. CA measures whether the output provides sufficient supported information for a reference-consistent clinical interpretation: 0, uninterpretable; 1, misleading; 2, usable only after major correction; 3, useful for a broad interpretation but missing important context; 4, adequate with minor omissions; 5, clear, sufficiently specific, and fully supported. Fabricated findings or unsupported recommenda- tions that could materially change clinical interpretation or management cap CA at 1. Recommen- dations are not required when the references provide no basis for them. Score normalization and aggregation. For each evaluated case i, the judge returns raw scores r PDA i ,r CA i ∈ 0, 1, 2, 3, 4, 5 and r KFP i ∈ [0, 1]. We linearly normalize PDA and CA by dividing their raw scores by 5, while retaining KFP on its original scale. All three metrics are therefore reported on a common [0, 1] scale, with higher values indicating better diagnostic fidelity. PDA represents a normalized ordinal rating of diagnostic correctness. Let N eval denote the number of evaluated cases in a dataset. The dataset-level scores are computed as PDA = 1 N eval N eval X i=1 r PDA i 5 , KFP = 1 N eval N eval X i=1 r KFP i , CA = 1 N eval N eval X i=1 r CA i 5 . (29) The Diagnostic Fidelity Avg. reported in the result tables is the unweighted mean of these three normalized metrics: Avg. DF = PDA + KFP + CA 3 .(30) CRESULT DETAILS C.1DATASET (1) PETCT, our shorthand for PETWB-REP (Xue et al., 2026), comprises 490 whole-body 18 F- FDG PET/CT examinations, each accompanied by de-identified bilingual imaging reports in Chi- nese and English and structured clinical information. The dataset covers various malignancies, including lung, liver, breast, prostate, and ovarian cancers. The reports describe both CT morpho- logical findings and PET metabolic findings, including lesion location, size, density, FDG uptake, SUV/SUVmax, lymph-node involvement, and distant metastases. We used all 490 English-language reports, of which 441 were assigned to the training set and 49 to the test set. (2) ViMed is a whole-body PET/CT dataset collected at a tertiary hospital in Vietnam. The official dataset contains PET/CT studies and their corresponding complete Vietnamese clinical reports collected in 2017, 2018, 2019, and 2023. The reports are organized into head and neck, chest, and abdominopelvic sections and include the examination indication, clinical history, demographic information, regional imaging findings, and diagnostic summary. We used the publicly available English version, in which the Vietnamese reports were translated into English using Gemma-4-26B-A4B-it. After excluding reports with empty imaging findings, 2,761 reports remained. We then performed a stratified split by year and month, yielding 2,484 training reports and 277 test reports. (3) MERLIN is a large- scale abdominal CT dataset covering normal and abnormal findings involving the hepatobiliary system, spleen, pancreas, kidneys, gastrointestinal tract, vasculature, lymph nodes, and other ab- dominopelvic structures. We selected 4,080 reports containing valid content in both the FINDINGS and SUMMARY sections. The main text of the FINDINGS section was used as the imaging de- scription, whereas the SUMMARY section was retained as the diagnostic reference. For the current 24 experiments, the reports were divided into 3,672 training reports and 408 test reports. (4) BDMAP contains abdominal CT volumes, paired imaging reports, and voxel-level annotations of structures such as liver, kidney, and pancreatic tumors. The reports were generated by RadGPT from organ and tumor segmentation results and contain quantitative and qualitative information, including organ volumes, CT attenuation values, lesion sizes, tumor burden, and organ status. After quality filtering and narrative-text deduplication, we randomly selected 2,000 reports to construct a BDMAP-derived text subset. Of these, 1,800 reports were used for training and 200 for testing. C.2BATCHED TEMPLATE CONSTRUCTION FOR LLM BASELINES Due to the context-length limitation of LLMs, the complete training corpus cannot generally be processed within a single model context. Therefore, for LLM baselines in the template construction task, a two-stage procedure is adopted. The complete training corpus is first divided into multiple smaller batches. In the first stage, each batch of reports is summarized into a local template. In the second stage, these local templates are consolidated while preserving their combined clinical coverage as much as possible. Batching Strategy. Let D train = x i N i=1 denote the training report set. A fixed random seed is first used to shuffle the reports, which are then partitioned into disjoint batches: D train = B ] b=1 B b .(31) The same batch partitionB b is reused across all LLM baselines to ensure a fair comparison. Since report lengths vary substantially across the four datasets, the number of reports per batch is deter- mined separately for each dataset so that the serialized input payloads remain approximately com- parable. Specifically, the batch sizes are set to 20, 80, 40, and 150 for PETCT, ViMed, MERLIN, and BDMAP, respectively. If the final remainder contains no more than half of the nominal batch size, it is merged into the preceding batch; otherwise, it is retained as a separate batch. Batch-Level Template Induction. For each batchB b , model f θ receives the same system instruc- tion, a dataset-specific modality description, a shared one-shot template example, and the report descriptions within the batch, producing a local template T (θ) b . Each template is represented as a JSON object organized by anatomical regions and organs. Gener- ated outputs undergo strict JSON parsing and schema validation. When strict parsing fails, the con- servative fallback described below attempts to recover a complete template object already present in the output. Only parsed or recovered objects that satisfy the template schema are admitted to the merging stage; outputs without such an object remain failures. Local Template Merging. All valid local templates are serialized and passed again to the same model. A merging strategyM θ is then used to consolidate them into the final template: T (θ) =M θ T (θ) 1 ,...,T (θ) B .(32) During this process, the original anatomical organization and clinically distinct information are pre- served. This procedure separates evidence acquisition from global normalization: the induction stage sum- marizes the information contained in each batch, whereas the merging stage reconciles naming and semantic granularity across batches. When the serialized local templates still exceed the context-length or generation budget of a single model call, the same merging operator is applied hierarchically. At level ℓ, the current templates are divided into groups G (ℓ) j j , and each group is reduced inde- pendently: T (ℓ+1) j =M θ n T (ℓ) i : i∈G (ℓ) j o .(33) This process is repeated until a single unified template remains. If the final merge group is too small, it is merged into the preceding group so that the intermediate templates summarize approximately comparable amounts of information. 25 Table 5: Performance comparison across PETCT, ViMed, MERLIN, and BDMAP in template cov- erage, information fidelity, and diagnostic fidelity, reported as unweighted macro-averages across the four datasets. Higher is better for all metrics. Among individual methods, best in bold and runner-up in underlined . Case = case-macro coverage; Key = key-micro coverage; Avg. = average; R-1/R-2/R-L = ROUGE-1/ROUGE-2/ROUGE-L; chrF = character n-gram F-score; chrF++ = char- acter and word n-gram F-score; BERT = BERTScore; PDA = Primary Diagnosis Accuracy; KFP = Key Finding Preservation; CA = Clinical Actionability. Template CoverageInformation FidelityDiagnostic Fidelity MethodCaseKeyAvg.R-1R-2R-LchrF chrF++ BERTAvg.PDAKFPCAAvg. Medical LLM baselines lingshu-32b0.5213 0.5339 0.5276 0.5171 0.2750 0.3669 0.39020.3742 0.8770 0.4667 0.4861 0.5172 0.5304 0.5112 medgemma-27b0.5388 0.5548 0.5468 0.6382 0.4104 0.4717 0.48730.4762 0.8970 0.5634 0.5710 0.6328 0.6232 0.6090 clinfusion-32b0.4894 0.4930 0.4912 0.3417 0.1849 0.2407 0.21800.2125 0.7995 0.3329 0.6236 0.6699 0.6873 0.6603 Medical LLM Avg.0.51650.52720.52190.49900.29010.35980.36510.35430.85780.45430.56020.60660.61360.5935 General LLM baselines deepseek-v4-flash0.6442 0.6626 0.6534 0.6228 0.3494 0.4326 0.46940.4571 0.8934 0.5374 0.6667 0.7099 0.7248 0.7004 deepseek-v4-pro0.6419 0.6478 0.6449 0.6130 0.3502 0.4401 0.49420.4759 0.8905 0.5440 0.6288 0.6779 0.6861 0.6643 qwen3.6-flash0.5866 0.6052 0.5959 0.5217 0.2970 0.3599 0.40190.3844 0.8800 0.4741 0.5403 0.5839 0.5887 0.5709 gemini-3.6-flash0.6362 0.6346 0.6354 0.6672 0.45810.51460.51360.5028 0.9007 0.59280.6140 0.6794 0.6661 0.6532 gpt-5.40.5877 0.6160 0.6019 0.6145 0.4222 0.4634 0.47420.4625 0.8985 0.5559 0.4883 0.5404 0.5397 0.5228 gemini-3.5-flash-lite 0.6442 0.6606 0.6524 0.6468 0.4396 0.4965 0.51710.5058 0.9015 0.5846 0.6492 0.6993 0.7069 0.6851 General LLM Avg.0.62350.63780.63060.61430.38610.45120.47840.46470.89410.54810.59790.64850.65200.6328 Template induction baseline ASTAR0.79380.82490.80930.6503 0.3530 0.4063 0.55180.52830.8909 0.5634 0.75810.80350.79430.7853 Ours0.83610.85130.84370.85850.72010.67910.80250.79380.93910.79880.81460.85330.83140.8331 C.3AGGREGATE RESULTS ACROSS DATASETS The unweighted macro-average results across PETCT, ViMed, MERLIN, and BDMAP (Table 5) complement the dataset-specific comparisons (Tables 1 and 2). Overall, ASTAR provides the strongest baseline performance when considering Template Coverage, Information Fidelity, and Diagnostic Fidelity jointly, although Gemini-3.6-Flash achieves the highest baseline score on In- formation Fidelity alone. Our method achieves the best overall performance across all three criteria, with aggregate scores of 0.8437, 0.7988, and 0.8331, respectively, and obtains the highest score on every individual metric reported in the table. C.4ABLATION ANALYSIS OF REVISION AND NOTE. To further disentangle the contributions of Revision and Note to template quality and clinical util- ity, we compare ASTAR with three sequential variants of our framework: Ours w/o (Note + Rev.), in which both Revision and Note are removed; Ours w/o Note, which retains Revision but removes Note; and the full model, Ours, which incorporates both components. Because Note influences template usage by supplementing guidance on observation priorities, abnor- mal manifestations, interpretation criteria, and slot completion rather than by introducing additional structural fields, we focus on 13 evaluation dimensions that are closely aligned with its functional pathway. These comprise eight Template Design dimensions (TD1, TD2, TD7, TD8, TD9, TD11, TD12, and TD13) and five Clinical Utility dimensions (CU1, CU2, CU4, CU5, and CU6). Col- lectively, these dimensions assess coverage of key structures, representation of abnormal findings, omission prevention, prevention of erroneous normal-state completion, uncertainty representation, clinical feasibility, sign-off safety, overall usability, reduction of missing information, reporting con- sistency, support for not-assessed states, communication efficiency, and overall satisfaction. For each dataset, we use an LLM judge to conduct exhaustive pairwise comparisons among the four candidate templates: ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and Ours. The four templates form six unordered pairs. Each pair is evaluated twice with the A/B presen- tation order reversed, yielding 12 directed comparisons in total to mitigate potential positional bias. In each comparison, the judge is provided with the complete slot schemas of both templates and, for each of the 13 dimensions, outputs a winner (A, B, or tie) together with an advantage magnitude m∈0, 1, 2, 3, where 0 denotes no distinguishable difference and 3 denotes a decisive advantage. 26 For dimension d, the directional advantage of the first-presented template over the second is defined as: a d (A,B) = m/3,if A is preferred, −m/3, if B is preferred, 0,if the judgment is a tie. (34) For a pair of templates X and Y , the two presentation orders are then symmetrized as: s d (X,Y ) = a d (X,Y )− a d (Y,X) 2 ,(35) so that judgments from both A/B orderings jointly determine the relative advantage of X over Y on dimension d. Because each template is compared against the other three candidates, its score on dimension d is computed as: S d (X) = 0.5 + 0.5× 1 3 X Y ̸=X s d (X,Y ),(36) which maps the resulting relative-preference score to the interval [0, 1]. A score of 0.5 indicates that the template is, on average, comparable to the other candidates in the current comparison set; values approaching 1 indicate consistent preference over competing templates, whereas values approaching 0 indicate consistent disadvantage. Finally, Selected Design is obtained by averaging S d (X) over the eight selected Template Design dimensions, while Selected Utility is obtained by averaging over the five selected Clinical Utility dimensions. Selected Overall is computed directly as the mean over all 13 selected dimensions. These scores therefore quantify the relative preference of each template within the current candidate set rather than an absolute assessment of clinical quality independent of the competing templates. Results. Across all four datasets, Revision and Note provide consistent cumulative improvements over the corresponding ablated variants (Figs. 6–13). Adding Revision increases the Selected Over- all score from 0.408 to 0.504 on PETCT, from 0.383 to 0.487 on ViMed, and from 0.355 to 0.609 on MERLIN, with a smaller increase from 0.446 to 0.476 on BDMAP. Incorporating Note yields a further and more pronounced improvement, raising the corresponding scores to 0.844, 0.844, 0.863, and 0.911, respectively. The dimension-wise results exhibit the same general pattern, with the full model consistently receiving stronger preferences across the selected Design Quality and Clinical Utility criteria. Overall, these results support complementary roles for Revision and Note, with Note contributing particularly strongly to the dimensions associated with clinically grounded tem- plate guidance and downstream usability. 27 TD1 TD2 TD7 TD8 TD9 TD11 TD12 TD13 (a) Selected Design dimensions 0.00.20.40.60.81.0 Dimension score CU1 CU2 CU4 CU5 CU6 (b) Selected Utility dimensions Performance across the 13 selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 6: Dimension-wise ablation analysis on the PETCT dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, compar- ing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension. Selected Design (8 dimensions) Selected Utility (5 dimensions) Selected Overall (13 dimensions) 0.0 0.2 0.4 0.6 0.8 1.0 Mean score 0.236 0.256 0.244 0.420 0.389 0.408 0.500 0.511 0.504 0.844 0.845 0.844 Mean scores over the selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 7: Aggregated ablation results on the PETCT dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. 28 TD1 TD2 TD7 TD8 TD9 TD11 TD12 TD13 (a) Selected Design dimensions 0.00.20.40.60.81.0 Dimension score CU1 CU2 CU4 CU5 CU6 (b) Selected Utility dimensions ViMed: Performance across the 13 selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 8: Dimension-wise ablation analysis on the ViMed dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, compar- ing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension. Selected Design (8 dimensions) Selected Utility (5 dimensions) Selected Overall (13 dimensions) 0.0 0.2 0.4 0.6 0.8 1.0 Mean score 0.319 0.233 0.286 0.372 0.400 0.383 0.469 0.517 0.487 0.841 0.850 0.844 ViMed: Mean scores over the selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 9:Aggregated ablation results on the ViMed dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. 29 TD1 TD2 TD7 TD8 TD9 TD11 TD12 TD13 (a) Selected Design dimensions 0.00.20.40.60.81.0 Dimension score CU1 CU2 CU4 CU5 CU6 (b) Selected Utility dimensions Merlin: Performance across the 13 selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 10:Dimension-wise ablation analysis on the MERLIN dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension. Selected Design (8 dimensions) Selected Utility (5 dimensions) Selected Overall (13 dimensions) 0.0 0.2 0.4 0.6 0.8 1.0 Mean score 0.156 0.200 0.173 0.372 0.328 0.355 0.607 0.611 0.609 0.865 0.861 0.863 Merlin: Mean scores over the selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 11: Aggregated ablation results on the MERLIN dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. 30 TD1 TD2 TD7 TD8 TD9 TD11 TD12 TD13 (a) Selected Design dimensions 0.00.20.40.60.81.0 Dimension score CU1 CU2 CU4 CU5 CU6 (b) Selected Utility dimensions Performance across the 13 selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 12:Dimension-wise ablation analysis on the BDMAP dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension. Selected Design (8 dimensions) Selected Utility (5 dimensions) Selected Overall (13 dimensions) 0.0 0.2 0.4 0.6 0.8 1.0 Mean score 0.184 0.139 0.167 0.444 0.450 0.446 0.465 0.494 0.476 0.906 0.917 0.911 Mean scores over the selected dimensions ASTAROurs w/o (Note + Rev.)Ours w/o NoteOurs Figure 13: Aggregated ablation results on the BDMAP dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. 31 C.5DETAILED TEMPLATE QUALITY RESULTS The overall template-quality results compare our method and all baselines across the four datasets under both expert and LLM evaluation (Fig. 3). Beyond the aggregate template-quality scores re- ported in the main figure, this appendix presents dimension-level design-quality and clinical-utility ratings and their aggregates for PETCT (Figs. 14–16), ViMed (Figs. 17–19), MERLIN (Figs. 20– 22), and BDMAP (Figs. 23–25). DQ1DQ2DQ3DQ4DQ5DQ6DQ7DQ8DQ9DQ10DQ11DQ12DQ13 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 3432323233222 3432323223222 3342313112222 3332323112222 3342323113222 3343423122223 3343323112323 2232322112212 2333423112222 3332323112222 3343423112323 Design quality across 13 dimensions 1 2 3 4 Ordinal score Figure 14:Design Quality evaluation on the PETCT dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3) of Template Quality. Darker colors indicate higher ordinal scores. 32 CU1CU2CU3CU4CU5CU6 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 333222 323222 324122 323122 334122 334122 333122 233122 333122 323122 343122 Clinical utility across six dimensions 1 2 3 4 Ordinal score Figure 15:Clinical Utility evaluation on the PETCT dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4) of Template Quality. Darker colors indicate higher ordinal scores. 33 1.82.02.22.42.62.8 Reported mean score Gemini-3.6-Flash DeepSeek-V4-Pro Lingshu-32B DeepSeek-V4-Flash GPT-5.4 Qwen3.6-Flash ASTAR Gemini-3.5-Flash-Lite MedGemma-27B ClinFusion-32B Ours 1.92 2.23 2.23 2.23 2.31 2.54 2.54 2.38 2.62 2.62 2.62 2.17 2.17 2.17 2.33 2.33 2.33 2.33 2.50 2.50 2.50 2.50 2.045 2.200 2.200 2.280 2.320 2.435 2.435 2.440 2.560 2.560 2.560 Mean design-quality and clinical-utility scores DQ1–DQ13 meanCU1–CU6 meanCombined mean Figure 16:LLM-as-judge Template Quality evaluation on the PETCT dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores. 34 DQ1DQ2DQ3DQ4DQ5DQ6DQ7DQ8DQ9DQ10DQ11DQ12DQ13 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 3432323133222 3422212233212 3332423112222 2332312112212 2332323112222 2232312112212 2332322122212 2232212112212 2232412113212 2232312112212 2332322112212 Design quality across 13 dimensions 1 2 3 4 Ordinal score Figure 17: Design Quality evaluation on the ViMed dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3) of Template Quality. Darker colors indicate higher ordinal scores. 35 CU1CU2CU3CU4CU5CU6 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 333122 322222 333122 223122 333122 223122 223122 223122 223122 223122 333122 Clinical utility across six dimensions 1 2 3 4 Ordinal score Figure 18: Clinical Utility evaluation on the ViMed dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4) of Template Quality. Darker colors indicate higher ordinal scores. 36 1.01.52.02.53.0 Mean ordinal score Gemini-3.6-Flash MedGemma-27B Lingshu-32B DeepSeek-V4-Pro GPT-5.4 Qwen3.6-Flash ClinFusion-32B ASTAR Gemini-3.5-Flash-Lite DeepSeek-V4-Flash Ours 1.77 1.85 1.85 1.92 2.00 2.08 2.00 2.23 2.15 2.31 2.54 2.00 2.00 2.00 2.00 2.00 2.00 2.33 2.17 2.33 2.33 2.33 1.885 1.925 1.925 1.960 2.000 2.040 2.165 2.200 2.240 2.320 2.435 Mean design-quality and clinical-utility scores DQ1–DQ13 mean CU1–CU6 mean Overall mean Figure 19:LLM-as-judge Template Quality evaluation on the ViMed dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores. 37 DQ1DQ2DQ3DQ4DQ5DQ6DQ7DQ8DQ9DQ10DQ11DQ12DQ13 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 3422223243122 2321212132112 1222212111212 1232312131111 1222312112212 1222212112111 1322322122212 1222312112212 1232322111111 2321212112212 1222212111111 Design quality across 13 dimensions 1 2 3 4 Ordinal score Figure 20: Design Quality evaluation on the MERLIN dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3) of Template Quality. Darker colors indicate higher ordinal scores. 38 CU1CU2CU3CU4CU5CU6 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 322222 221111 221122 223121 222122 222111 222122 222122 222111 222121 221111 Clinical utility across six dimensions 1 2 3 4 Ordinal score Figure 21:Clinical Utility evaluation on the MERLIN dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4) of Template Quality. Darker colors indicate higher ordinal scores. 39 1.01.52.02.53.0 Mean ordinal score ClinFusion-32B MedGemma-27B ASTAR GPT-5.4 DeepSeek-V4-Flash Lingshu-32B Gemini-3.5-Flash-Lite Gemini-3.6-Flash DeepSeek-V4-Pro Qwen3.6-Flash Ours 1.38 1.46 1.77 1.62 1.54 1.69 1.69 1.69 1.69 1.92 2.46 1.33 1.50 1.33 1.50 1.67 1.67 1.83 1.83 1.83 1.83 2.17 1.355 1.480 1.550 1.560 1.605 1.680 1.760 1.760 1.760 1.875 2.315 Mean design-quality and clinical-utility scores DQ1–DQ13 mean CU1–CU6 mean Overall mean Figure 22: LLM-as-judge Template Quality evaluation on the MERLIN dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores. 40 DQ1DQ2DQ3DQ4DQ5DQ6DQ7DQ8DQ9DQ10DQ11DQ12DQ13 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 3432323223222 3322212123212 2222312112212 2122322112212 1122312112211 1122311111111 1122211112111 1122211112111 1122211112111 1121211111111 1121211111111 Design quality across 13 dimensions 1 2 3 4 Ordinal score Figure 23:Design Quality evaluation on the BDMAP dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3) of Template Quality. Darker colors indicate higher ordinal scores. 41 CU1CU2CU3CU4CU5CU6 Evaluation dimension Ours ASTAR DeepSeek-V4-Flash DeepSeek-V4-Pro Gemini-3.5-Flash-Lite MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Lingshu-32B ClinFusion-32B Method 333222 222122 222122 222122 122111 121111 121111 122111 122111 121111 121111 Clinical utility across six dimensions 1 2 3 4 Ordinal score Figure 24:Clinical Utility evaluation on the BDMAP dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4) of Template Quality. Darker colors indicate higher ordinal scores. 42 1.01.52.02.53.0 Mean ordinal score Lingshu-32B ClinFusion-32B MedGemma-27B Qwen3.6-Flash Gemini-3.6-Flash GPT-5.4 Gemini-3.5-Flash-Lite DeepSeek-V4-Flash DeepSeek-V4-Pro ASTAR Ours 1.15 1.15 1.31 1.31 1.31 1.31 1.54 1.77 1.77 2.00 2.54 1.17 1.17 1.17 1.17 1.33 1.33 1.33 1.83 1.83 1.83 2.50 1.160 1.160 1.240 1.240 1.320 1.320 1.435 1.800 1.800 1.915 2.520 Mean design-quality and clinical-utility scores DQ1–DQ13 mean CU1–CU6 mean Overall mean Figure 25:LLM-as-judge Template Quality evaluation on the BDMAP dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores. 43 C.6FAILURE ANALYSIS OF ADDITIONAL BASELINES Beyond the baseline results reported in the main text, we evaluated several additional models on the same template-induction task. These baselines were not included in the main comparison because they failed during template construction or the subsequent merge stage. The observed failures arose from insufficient context capacity, unstable long-form generation, malformed structured outputs, and repetition degeneration. We report these results to document the scope of the comparison and to distinguish recoverable formatting errors from cases in which no usable template was generated. C.6.1ADDITIONAL BASELINE MODELS Doubao Seed series. We first removed endpoints intended for embedding, image, video, code, or role-playing applications, as well as models with a 32k-token context window. In the screening workload, the largest induction prompt contained 26,526 tokens and required approximately 13k additional tokens for the generated template, already exceeding a 32k-token window. We screened the remaining long-context candidates using a real induction prompt (Table 6). The requests did not specify max_tokens. Nevertheless, four models terminated with finish_reason=length after producing between 11,541 and 21,623 output tokens, which indicates that an output-length limit was reached; the observed finish reason alone does not distinguish a service default budget from a model limit. The two Doubao 2.1 endpoints did not return within an acceptable time budget, despite responding normally to a trivial JSON request. The apparently successful seed-1.6 result was not stable under repeated evaluation. With an identical prompt and temperature set to zero, two consecutive calls produced 10,099 and 10,098 output tokens, respectively. The first completed normally, whereas the second terminated with finish_reason=length and five unmatched braces. In the full induction run, only 6 of 22 batches (27%) succeeded without retrying. Seventeen retry rounds were required to obtain outputs for all 22 batches, after which the merge call was still truncated at 14,514 output tokens. By com- parison, successful merge outputs from the other models contained approximately 24k–44k tokens. We therefore exclude the Doubao models from the primary comparison because they cannot reliably execute the fixed protocol, rather than because of their downstream template quality. Local medical models. For local models, the configured context window provides an initial screen- ing criterion. We measured peak prompt and template lengths for each dataset to characterize context demand (Table 7). The sums of the separately reported input and output peaks range from 38,204 to 43,624 tokens; these sums are screening estimates rather than measurements of a single request. HuatuoGPT-Vision-34B has a 4,096-token window without RoPE scaling and cannot accommodate even the prompt skeleton plus one report and the required output bud- get. Hulu-Med-32B has an effective 32,768-token window. Although its configuration contains sliding_window=131072, sliding-window attention is disabled. The observed merge work- load also included PET/CT inputs exceeding 63k tokens and a MERLIN output of 65,067 tokens. Both models were excluded from the main comparison under the evaluated protocol. These obser- vations do not establish infeasibility for every alternative batching or hierarchical-merging schedule. The remaining long-context candidates were Lingshu-32B (128k), MedGemma-27B-IT (131,072), Hulu-Med-30A3 (262,144), and ClinFusion-32B (262,144). Hulu-Med-30A3 was subsequently excluded from the primary comparison. In two of its 22 raw outputs, the model remained inside an unclosed <think> block, entered a repetitive loop, and terminated without pro- ducing a single opening brace. This behavior is a semantic generation failure: no template object exists for a parser to recover. C.6.2FAILURE TAXONOMY FOR STRUCTURED OUTPUTS We analyzed 409 raw outputs from the local-model runs. The strict extraction path, which removes Markdown fences and reasoning blocks before applying ordinary JSON parsing, accepted 294 out- puts (71.9%). The remaining 115 outputs (28.1%) failed strict parsing. Parser exceptions alone were not a reliable description of the underlying failure. For example, a restart that occurs in the middle of a string is reported as an invalid control character, although the relevant event is that the model abandoned one JSON attempt and began another. We therefore classify failures by their generated structure rather than by the exception message. Truncation accounts for the majority of strict-parsing failures in this structural taxonomy (Table 8). 44 Table 6: Feasibility screening of the Doubao Seed series on a real induction prompt. No request- level output-token limit was specified. ModelFinish reasonOutput tokensOutcome seed-2.0-prolength21,623truncated JSON seed-1.8length18,836truncated JSON seed-2.0-minilength18,634truncated JSON seed-2.0-litelength11,541truncated JSON seed-1.6stop6,982valid JSON in screening seed-2.1-pro–no response after approximately 40 min seed-2.1-turbo–timeout at 180 s and 1,800 s Table 7: Observed peak input and output lengths for induction, measured with the Qwen tokenizer on completed runs. The final column sums the two peaks, which need not occur in the same request. DatasetPeak inputPeak outputSum of peaks PET/CT24,87213,33238,204 ViMed29,3878,96438,351 BDMAP39,4613,55843,019 MERLIN30,03913,58543,624 C.6.3FALLBACK RECOVERY When strict JSON parsing fails, we apply a conservative fallback that retains only complete struc- tures already present in the output. Depending on the failure type, it removes an incomplete suffix, discards extra closing delimiters, or separates multiple JSON attempts produced by a generation restart. The recovered object must still satisfy the template schema, and the procedure never intro- duces new slots, keys, or values. Outputs without an extractable template remain failures. Fallback recovery prevented 12 call-level failures, all in local-model runs, while the evaluated hosted-API outputs passed strict parsing. Without this step, a malformed batch output containing a recoverable template could be discarded before merging. The fallback therefore prevents usable outputs from being discarded because of local formatting errors without changing their semantic content. C.6.4REPETITION DEGENERATION AND CANONICALIZATION Structural recovery exposes a second failure mode: some models emit long sequences of repeated but individually well-formed components before truncation. In one MedGemma-27B-IT output, fallback recovered a slot whose candidate-value list Ω contained 1,719 entries but only six unique values. One 130-character value was repeated 1,714 times. Removing exact duplicates reduced this slot from 226,751 to 635 characters without deleting any unique value. Repetition was strongly concentrated in outputs that required fallback. The duplicate rate within Ω was 5.0% for strictly parsed local outputs and 80.7% for fallback-recovered outputs. This association suggests a common generation process: repetition expands the sequence, the expanded sequence reaches an output limit, and the resulting truncation leaves an invalid JSON suffix. The association does not by itself establish that repetition caused every truncation, but it explains why malformed outputs contain substantially more duplicated content than directly parsed outputs. We observed two distinct forms of degeneration. MedGemma-27B-IT, Lingshu-32B, and Hulu-Med-30A3 primarily repeated values within the same Ω list. ClinFusion-32B, in con- trast, repeated entire slots. A representative ClinFusion-32B output contained 509 recovered slots but only 32 unique slots, while each individual Ω list contained two non-duplicated values. We therefore canonicalize recovered templates at two levels: exact duplicate values are removed within each Ω, and slots with the same canonical key are merged while preserving the union of their unique values. Across 114 recovered templates, value-level deduplication removed 80.7% of the recovered characters, and slot-level merging increased the total reduction to 84.2%. The operation is content-preserving with respect to the set of unique generated keys and values. 45 Table 8: Structural causes of the 115 outputs rejected by strict JSON parsing. Bracket accounting ignores delimiters that occur inside strings. CauseCountStructural signature Truncation781–59 missing and up to 14 missing ] Generation restart17a second JSON fence or a second top-level desc Excess closing delimiter17an otherwise complete object followed by extra delimiters Balanced but invalid syntax3e.g., an unambiguous trailing comma DatasetExamplesCategoricalNumerical PETCT12111 ViMed12111 MERLIN14131 BDMAP1284 Total50437 Table 9: Distribution of representative examples from the final templates presented in this appendix. Together, these analyses explain why the additional baselines were absent from the main results and how malformed local-model outputs were handled. Models that could not complete template induction and merging were excluded, recoverable formatting errors were repaired without adding semantic content, and outputs containing no usable template remained explicit generation failures. DREPRESENTATIVE EXAMPLES FROM FINAL TEMPLATES The final templates are hierarchical clinical specifications organized by anatomical paths. We present 50 representative examples from the final templates constructed for PETCT, ViMed, MERLIN, and BDMAP (Table 9). These examples cover organ status, lesion morphology, metabolic activity, quan- titative measurements, temporal change, post-surgical findings, and anatomical relationships. Each example presents its anatomical path, canonical field key k, value type τ , allowed values Ω, and clinical note. All fields are taken directly from the corresponding final template (Tables 10–13). D.1PETCT The PETCT examples span the brain, thyroid, lungs, mediastinal lymph nodes, liver, kidneys, fe- male pelvis, and skeletal system. They show how morphology and FDG activity are represented by separate but anatomically organized template fields. PETCT template specification brain_fdg_metabolism_status categorical Anatomical path head_and_neck/brain Allowed values Ω [normal, increased, decreased, absent, asymmetric, focal_hypermetabolism] Clinical note Normal: symmetrical cortical & subcortical uptake; abnormal: focal hypermetabolism (tumor/inflammation), global hypometabolism (neurodegeneration), asymmetry, or absent uptake. 46 PETCT template specification (continued) thyroid_overall_status categorical Anatomical path head_and_neck/thyroid Allowed values Ω [normal, enlarged, atrophic, small, heterogeneous, irregular, asymmetric, nodular, diffuse, residual] Clinical note Normal: homogeneous, smooth, SUVmax < liver; abnormal: focal/diffuse uptake > liver, enlarged, heterogeneous, nodular, or irregular margins. lung_fdg_metabolism categorical Anatomical path chest_and_thorax/lungs Allowed values Ω [normal, increased, decreased, focal, diffuse, abnormal] Clinical note Normal: homogeneous, low-grade FDG uptake (SUVmax <2.5). Focal/diffuse/increased uptake suggests inflam- mation, infection, or malignancy. lung_fdg_suvmax numerical Anatomical path chest_and_thorax/lungs Allowed values Ω [<num>] Clinical note Normal lung parenchyma SUVmax ~0.5-1.0; suspicious lesion usually >2.5, but inflammatory uptake can overlap. Use visual comparison with liver blood pool. lung_lesion_margin categorical Anatomical path chest_and_thorax/lungs/lung_lesion_margin Allowed values Ω [clear, indistinct, spiculated, lobulated, smooth, irregular] Clinical note Smooth/lobulated margins often benign; spiculated/irregular margins suggest malignancy. Indistinct margin may indicate inflammation or infiltration. nodule_type categorical Anatomical path chest_and_thorax/lungs/nodules Allowed values Ω [solid, part_solid, ground_glass, calcified, plaque] Clinical note Solid: homogeneous opacity; part-solid: ground-glass + solid component; ground-glass: hazy, no solid; calcified: benign. Compare with prior CT. 47 PETCT template specification (continued) hilar_mediastinal_lymph_node_status categorical Anatomical path chest_and_thorax/mediastinum_and_hilar_lymph_nodes Allowed values Ω [normal, enlarged, absent, multiple] Clinical note Normal: no FDG-avid hilar/mediastinal nodes above background. Abnormal: focal uptake >= mediastinal blood pool, with or without enlargement (>=10m short axis). liver_focal_lesions categorical Anatomical path abdomen_and_pelvis/liver Allowed values Ω [absent, present, single, multiple, cystic, solid, mixed, low_density, high_density] Clinical note Normal: no focal uptake > adjacent liver background. Solid/mixed lesions with SUVmax >3.0 or >liver suggest malignancy. Cysts: typically no uptake. liver_fdg_metabolism categorical Anatomical path abdomen_and_pelvis/liver Allowed values Ω [normal, increased, decreased, focal, diffuse, absent, uneven] Clinical note Normal liver FDG uptake is homogeneous and similar to blood pool; suspicious if focal/diffuse increased uptake exceeds blood pool or is heterogeneous. kidney_morphology_and_size categorical Anatomical path abdomen_and_pelvis/kidneys_and_urinary_tract Allowed values Ω [normal, enlarged, atrophic, irregular, asymmetric] Clinical note Normal: smooth contour, <12 cm length; enlarged >12 cm (e.g., obstruction, mass); atrophic <8 cm (e.g., scarring); irregular/asymmetric suggests focal pathology. uterus_morphology_size_and_contour categorical Anatomical path abdomen_and_pelvis/uterus_and_cervix Allowed values Ω [normal, enlarged, atrophic, deformed, irregular, nodular, absent, not visualized] Clinical note Normal: anteverted, smooth contour, homogeneous; enlarged: >10 cm length or >5 cm AP diameter; abnormal: irregular contour, nodularity, or focal deformity suggests leiomyoma or malignancy. 48 PETCT template specification (continued) skeletal_fdg_metabolism_status categorical Anatomical path musculoskeletal_and_spine/systemic_bones Allowed values Ω [normal, abnormal, increased, decreased, absent] Clinical note Normal: homogeneous low-grade uptake ~SUVmax 1.5-2.5 (< liver); abnormal: focal or diffuse increased uptake, often SUVmax > liver, suggesting malignancy or inflammation. Table 10: Representative examples from the final PETCT template. D.2VIMED The ViMed examples emphasize physiological versus suspicious FDG patterns, nodal morphology and station labels, organ-level uptake, quantitative SUVmax, and post-surgical imaging findings. ViMed template specification head_and_neck_fdg_characteristics categorical Anatomical path head_and_neck Allowed values Ω [absent, physiological, mild, focal, diffuse, asymmetric, increased] Clinical note Normal: symmetric, mild FDG uptake in Waldeyer’s ring, salivary glands, and muscles; suspicious: focal, asymmetric, or intense uptake (SUVmax > 4-5) suggests malignancy. tonsil_fdg_uptake categorical Anatomical path head_and_neck/oral_cavity_and_oropharynx Allowed values Ω [absent, mild, moderate, severe] Clinical note Normal: symmetric, mild-to-moderate FDG uptake (SUVmax typically 2-4). Abnormal: marked asymmetry, focal intense uptake, or SUVmax >5-6, suspicious for malignancy or infection. cervical_nodal_level categorical Anatomical path head_and_neck/cervical_and_supraclavicular_lymph_nodes Allowed values Ω [iia, iib, i, iv, v, vi, iia_iib_i] Clinical note Normal: no FDG-avid lymph nodes in any cervical level. Suspicious: focal FDG uptake > liver activity, especially in levels I-IV, with irregular margins or size >1 cm short axis. 49 ViMed template specification (continued) lung_fdg_uptake_status categorical Anatomical path thorax_and_chest/lungs_and_pulmonary_parenchyma Allowed values Ω [absent, normal, increased, decreased, focal, diffuse, physiological] Clinical note Normal: homogeneous, diffuse low-grade uptake (SUVmax <2.5). Abnormal: focal or increased uptake >2.5 suggests malignancy; diffuse increased uptake may indicate inflammation. axillary_lymph_node_fdg_status categorical Anatomical path thorax_and_chest/axillary_lymph_nodes Allowed values Ω [absent, normal, mild, increased] Clinical note Normal: absent/low-level FDG uptake (similar to blood pool). Abnormal: focal/asymmetric increased uptake (SUVmax >2.5) suggests metastasis or reactive inflammation. liver_fdg_uptake_intensity categorical Anatomical path abdomen_and_pelvis/liver_and_intrahepatic_biliary_tree Allowed values Ω [normal, increased, reduced, absent] Clinical note Characterizes the overall intensity of hepatic FDG uptake. liver_fdg_uptake_pattern categorical Anatomical path abdomen_and_pelvis/liver_and_intrahepatic_biliary_tree Allowed values Ω [homogeneous, heterogeneous, focal, diffuse] Clinical note Characterizes the spatial distribution pattern of hepatic FDG uptake. colon_fdg_uptake_status categorical Anatomical path abdomen_and_pelvis/large_intestine_and_rectum Allowed values Ω [physiological, inflammatory, focal, diffuse, heterogeneous, increased, mildly_increased] Clinical note Normal: diffuse, mild, homogeneous uptake (grade 1-2); suspicious: focal, intense, or segmental uptake exceeding liver activity, or irregular/mass-like. 50 ViMed template specification (continued) spleen_fdg_uptake categorical Anatomical path abdomen_and_pelvis/spleen Allowed values Ω [physiological, pathological, increased, decreased, focal, diffuse, heterogeneous] Clinical note Normal spleen: diffuse, homogeneous uptake <= liver; pathological: focal, heterogeneous, or diffuse > liver uptake (e.g., lymphoma, metastasis). lymph_node_morphology categorical Anatomical path lymph_nodes Allowed values Ω [normal, enlarged, reactive, calcified, necrotic, preserved_fatty_hila, enhancing] Clinical note Normal: homogeneous, smooth margins, preserved fatty hila, SUVmax <= mediastinal blood pool; abnormal: focal/heterogeneous uptake, irregular margins, loss of fatty hilum, necrosis, calcification. fdg_suvmax numerical Anatomical path suvmax_measurements Allowed values Ω [<num>] Clinical note Normal SUVmax varies by organ; for liver ~2.0-3.0, blood pool ~1.2-2.0; lesion SUVmax > liver often suggests malignancy. surgical_site_imaging_findings categorical Anatomical path musculoskeletal_and_other/surgical_sites_and_postoperative_changes Allowed values Ω [absent, normal, mild, diffuse, diffuse_increased, increased, heterogeneous_density, fdg_avid, stable, new, recurrence, resected, soft_tissue_lesion, postoperative_inflammation] Clinical note Post-surgical site: normal = linear/mild SUV uptake (<= adjacent background); suspicious = increasing, heteroge- neous, or FDG-avid focal uptake (SUVmax > liver or time-dependent rise). Table 11: Representative examples from the final ViMed template. D.3MERLIN The MERLIN examples cover gastrointestinal caliber and obstruction, hepatobiliary findings, peri- toneal abnormalities, renal morphology and function, vascular patency, osseous lesions, post- surgical change, longitudinal comparison, and quantitative measurement. 51 MERLIN template specification gastric_intestinal_caliber categorical Anatomical path gastrointestinal/gastric_intestinal_caliber Allowed values Ω [normal, distended, collapsed, narrowed, absent] Clinical note Normal: thin-walled, minimally distended; abnormal: marked dilation (>6 cm small bowel, >9 cm colon), wall thickening, or abrupt transition indicating obstruction. small_bowel_dilation_status categorical Anatomical path gastrointestinal/small_bowel Allowed values Ω [normal, dilated, focally_dilated, collapsed, decompressed, narrow] Clinical note Normal caliber is <=3 cm; dilation >3 cm with proximal distension indicates obstruction, while collapse denotes <2 cm. biliary_dilation_severity categorical Anatomical path hepatobiliary/bile_ducts Allowed values Ω [absent, mild, moderate, severe] Clinical note Normal CBD <=6 m; mild 7-10 m, moderate 11-15 m, severe >15 m. Intrahepatic ducts normally <1/3 adjacent portal vein diameter. lesion_characterization categorical Anatomical path hepatobiliary/liver Allowed values Ω [cyst, hemangioma, metastasis, HCC, cholangiocarcinoma, abscess, granuloma_calcification, neuroendocrine_tumor, benign, malignant, indeterminate, too_small_to_characterize] Clinical note Normal: homogeneous parenchyma. Characterize: cysts non-enhancing water-density; hemangiomas peripheral nodular fill-in; HCC/metastases arterial hyperenhancement; abscesses rim enhancement. pneumoperitoneum_free_air categorical Anatomical path peritoneum_mesenteric_and_abdominal_wall/peritoneum Allowed values Ω [absent, trace, small, moderate, large, multiple_foci, loculated] Clinical note Normal: absent. Abnormal: -1000 HU gas anterior to liver/spleen or outlining bowel; loculation or large volume indicates perforation or contained leak. 52 MERLIN template specification (continued) ascites_free_fluid categorical Anatomical path peritoneum_mesenteric_and_abdominal_wall/peritoneum Allowed values Ω [absent, trace, small, moderate, large, simple_fluid] Clinical note Normal: absent; trace/small shows water-density fluid (~0-20 HU) in dependent spaces, while moderate/large extends superiorly separating bowel and appears homogeneous if simple. kidney_morphology_and_appearance categorical Anatomical path genitourinary_and_adrenal/kidneys Allowed values Ω [normal, abnormal, lobulated, irregular, homogeneous, heterogeneous, symmetric] Clinical note Normal kidneys show smooth contours and homogeneous enhancement; abnormal morphology features irregular margins, heterogeneous attenuation, or asymmetry. nephrolithiasis_status categorical Anatomical path genitourinary_and_adrenal/kidneys Allowed values Ω [absent, present, obstructing, nonobstructing, punctate, single, multiple, calcific_density] Clinical note Normal: absent. Present: hyperdense (>100 HU) collecting system foci; obstructing causes hydronephrosis and perinephric stranding, whereas nonobstructing shows no dilation. kidney_excretion categorical Anatomical path genitourinary_and_adrenal/kidneys Allowed values Ω [normal, delayed, absent, symmetric, prompt, compromised, retained_contrast] Clinical note Normal: prompt symmetric opacification of collecting systems and ureters; abnormal: delayed/asymmetric excre- tion or retained parenchymal contrast suggests obstruction or impaired function. vessel_patency categorical Anatomical path vascular_and_lymph_nodes/vessels Allowed values Ω [patent, partially_occluded, occluded, attenuated] Clinical note Normal: continuous homogeneous contrast opacification without filling defects; abnormal: partial occlusion shows focal luminal narrowing or intraluminal thrombus, occlusion shows absent opacification. 53 MERLIN template specification (continued) bone_lesion_status categorical Anatomical path musculoskeletal_and_spine/bones Allowed values Ω [absent, present, diffuse, multifocal, progressive, contusion, erosion, sclerosis] Clinical note Normal: intact cortex, uniform trabeculae; suspicious: focal lytic (hypodense) or sclerotic (hyperdense) lesions, cortical breach, periosteal reaction, or soft-tissue mass. postsurgical_changes categorical Anatomical path other/postsurgical_changes Allowed values Ω [absent, present, normal, stable, unchanged, organized_fluid_collection] Clinical note Normal: surgical clips, mild fat stranding, expected resection defect; suspicious: new/enlarging rim-enhancing fluid, abscess, recurrent mass, or obstruction. temporal_stability_2 categorical Anatomical path other/temporal_stability_and_change Allowed values Ω [stable, improved, progressed, new, resolved] Clinical note Compare to prior CT. Stable: <20% size change, unchanged enhancement. Progressed: >=20% growth or new lesions. Improved/Resolved: decreased size or resolution. size_measurement numerical Anatomical path measurements Allowed values Ω [<num> cm] Clinical note Size interpretation is anatomy-specific; e.g., lymph nodes >1 cm short axis denote adenopathy, while solid organ lesions require correlation with attenuation and enhancement. Table 12: Representative examples from the final MERLIN template. D.4BDMAP The BDMAP examples focus on liver, spleen, pancreas, kidneys, and peripancreatic vessels. They highlight CT attenuation, morphology, lesion measurement, enhancement, and adjacent-structure relations. 54 BDMAP template specification liver_status categorical Anatomical path liver_and_biliary/liver Allowed values Ω [normal, abnormal] Clinical note Normal: homogeneous attenuation, smooth contour, uniform enhancement; Abnormal: heterogeneous density, nodular margins, focal hypo/hyperattenuating lesions, or altered enhancement. liver_attenuation_hu numerical Anatomical path liver_and_biliary/liver Allowed values Ω [<num> HU] Clinical note Normal hepatic attenuation is 40-60 HU; values below 40 HU indicate steatosis, whereas above 60-70 HU suggest iron overload. liver_attenuation categorical Anatomical path liver_and_biliary/liver Allowed values Ω [normal, decreased, increased, steatosis, heterogeneous] Clinical note Normal parenchyma is homogeneous (~50-60 HU); decreased/steatosis appears hypodense vs spleen (<40 HU), increased suggests iron/glycogen (>70 HU), and heterogeneous implies cirrhosis. spleen_density numerical Anatomical path spleen/splenic_parenchyma Allowed values Ω [<num> HU] Clinical note Normal spleen measures 40-60 HU on portal venous phase; focal or diffuse attenuation outside this range suggests cysts, infarcts, or infiltrative disease. spleen_parenchyma_appearance categorical Anatomical path spleen/splenic_parenchyma Allowed values Ω [normal, homogeneous, heterogeneous, coarse, abnormal, decreased, increased, splenomegaly, hypersplenism] Clinical note Normal: homogeneous attenuation, smooth margins; abnormal: coarse/heterogeneous texture, focal hypoattenu- ating lesions, or splenomegaly (>13 cm). 55 BDMAP template specification (continued) pancreas_normality categorical Anatomical path pancreas_and_peripancreatic_vessels/pancreas Allowed values Ω [normal, unremarkable, within normal limits] Clinical note Normal: homogeneous enhancement, smooth margins, non-dilated duct; abnormal: focal hypoattenuating mass, irregular borders, ductal dilatation, or peripancreatic stranding. pancreas_size_morphology categorical Anatomical path pancreas_and_peripancreatic_vessels/pancreas Allowed values Ω [normal, enlarged, mildly_enlarged, focal_enlargement, atrophic] Clinical note Normal: homogeneous parenchyma, smooth contour, body <=3 cm; abnormal: diffuse enlargement >3 cm, focal mass, or atrophy with lobularity and fat infiltration. pancreas_enhancement_pattern categorical Anatomical path pancreas_and_peripancreatic_vessels/pancreas Allowed values Ω [normal, homogeneous, heterogeneous, hypoenhancing, hyperattenuating, hypoattenuating, isoattenuating, calcified] Clinical note Normal shows homogeneous enhancement; focal hypoattenuation/hypoenhancement suggests carcinoma, while heterogeneity, calcification, or hyperenhancement indicate chronic pancreatitis or tumors. pancreas_mean_hu numerical Anatomical path pancreas_and_peripancreatic_vessels/pancreas/pancreas_mean_hu Allowed values Ω [<num> HU] Clinical note Normal parenchyma measures 30-50 HU non-contrast; deviations below 30 HU suggest steatosis, while elevations above 50 HU indicate fibrosis or inflammation. pancreas_lesion_size numerical Anatomical path pancreas_and_peripancreatic_vessels/pancreatic_lesions Allowed values Ω [<num> cm] Clinical note Normal: no focal mass; abnormal: solid lesion >=2 cm or cyst >=3 cm typically warrants further evaluation. 56 BDMAP template specification (continued) renal_lesion_attenuation categorical Anatomical path kidneys_and_adrenals/kidneys/kidney_lesion_attenuation Allowed values Ω [hypoattenuating, hyperattenuating, isoattenuating, normal, mixed, varying] Clinical note Hypoattenuating suggests simple cysts (<10 HU); hyperattenuating indicates hemorrhage or calcification (>20 HU), while isoattenuating blends with cortex and mixed denotes complex solid components. vascular_involvement categorical Anatomical path pancreas_and_peripancreatic_vessels/peripancreatic_vessels/vascular_involvement Allowed values Ω [absent, present, contact, encased] Clinical note Normal: intact fat plane, smooth vessel wall; abnormal: fat plane loss, luminal narrowing, irregular contour, or circumferential contact indicates invasion. Table 13: Representative examples from the final BDMAP template. EREPRESENTATIVE EXAMPLES FROM KNOWLEDGE GRAPH The PETCT examples illustrate anatomical organization and finding–diagnosis relations (Figs. 26 and 27). The ViMed examples show the corresponding anatomical and diagnostic structures (Figs. 28 and 29). The MERLIN examples pair anatomical relations with report-evidence links to clinical-significance outcomes (Figs. 30 and 31). The BDMAP examples show anatomical relations and diagnostic or clinical-status targets (Figs. 32 and 33). 57 real parent entity anatomical entity structure / location measurement / attribute finding / status association / inference additional real triples; labels omitted hepatic fissure FDG uptake liver shape liver size liver parenchyma FDG uptake normal morphology main pancreatic duct pancreas FDG uptake pancreatic shape gallbladder wall gallbladder size gallbladder FDG uptake FDG uptake gallbladder shape size spleen FDG uptake spleen size FDG uptake density absence of widening SUVmax abdomen FDG uptake portal vein dilation FDG uptake left left kidney FDG uptake renal parenchyma left ureter cystic lesion liver pancreas gallbladder spleen portal vein left kidney upper abdomen PETCT · Anatomical KG Figure 26: Visualization of the PETCT anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure. 58 PETCT · Diagnostic KG supports diagnosis metastatic pattern benign pattern rules out visual grouping (not a KG triple) canonical evidence; size = support PETCT canonical concept graph lung cancer hypermetabolic pulmonary lesion hypermetabolic lymphadenopathy hypermetabolic soft tissue lesion hypermetabolic pulmonary lesion hypermetabolic renal lesion non hypermetabolic pulmonary finding reactive lymphadenopathy hypermetabolic lymphadenopathy non hypermetabolic lymph nodes breast cancer hypermetabolic breast lesion hypermetabolic lymphadenopathy metastatic osseous lesion hypermetabolic hepatic lesion hypermetabolic pulmonary lesion non hypermetabolic breast finding cervical cancer hypermetabolic cervical lesion hypermetabolic lymphadenopathy metastatic soft tissue lesion hypermetabolic soft tissue lesion hypermetabolic soft tissue lesion inflammatory pulmonary nodules inflammatory pulmonary finding non hypermetabolic pulmonary finding pancreatic cancer hypermetabolic pancreatic lesion hypermetabolic lymphadenopathy metastatic hepatic lesion hypermetabolic pulmonary lesion metastatic peritoneal lesion osseous imaging finding Figure 27:Concept-level diagnostic knowledge graph for PETCT. Representative canonical imaging-evidence concepts are connected to diagnostic targets through the extracted diagnostic re- lation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and diagnostic target. Gray dashed links indicate visual grouping and are not knowledge-graph edges. 59 real parent entity anatomical entity structure / location measurement / attribute finding / status association / inference additional real triples; labels omitted physiological FDG uptake fatty liver liver left lobe enlarged physiological uptake physiological uptake abdomen and pelvis physiological FDG uptake normal spleen size physiological tracer uptake absent FDG uptake lymph node size abdomen and pelvis node count node size physiological FDG uptake abdomen and pelvis colon status colon dilation increased FDG metabolism physiological FDG uptake abdomen and pelvis adrenal SUVmax physiological uptake adrenal physiological uptake increased FDG metabolism perirectal fat infiltration mucosal thickening FDG avidity perirectal fat liver spleen inguinal lymph nodes colon adrenals rectum abdomen and pelvis ViMed · Anatomical KG Figure 28: Visualization of the ViMed anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure. 60 ViMed · Diagnostic KG supports diagnosis metastatic pattern benign pattern rules out visual grouping (not a KG triple) canonical evidence; size = support ViMed canonical concept graph lung cancer hypermetabolic pulmonary lesion hypermetabolic lymphadenopathy hypermetabolic osseous lesion physiologic pulmonary uptake esophageal carcinoma hypermetabolic esophageal lesion hypermetabolic lymphadenopathy nasopharyngeal carcinoma hypermetabolic lymphadenopathy hypermetabolic nasopharyngeal lesion reactive lymphadenopathy non hypermetabolic lymph nodes inflammatory soft tissue finding non hypermetabolic soft tissue finding breast cancer hypermetabolic breast lesion hypermetabolic lymphadenopathy gastric carcinoma hypermetabolic gastric lesion hypermetabolic peritoneal lesion colorectal carcinoma hypermetabolic soft tissue lesion metastatic pulmonary lesion inflammation inflammatory pulmonary finding soft tissue imaging finding cervical cancer metastatic soft tissue lesion physiologic cervical uptake liver cancer recurrence hepatic imaging finding Figure 29: Concept-level diagnostic knowledge graph for ViMed. The visualization connects representative canonical imaging-evidence concepts to diagnostic targets through the extracted di- agnostic relation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and diagnostic target. Gray dashed links indicate visual grouping and are not knowledge-graph edges. 61 real parent entity anatomical entity structure / location measurement / attribute finding / status association / inference additional real triples; labels omitted normal appearance present pancreatic duct pancreatic mass normal status spleen size normal spleen contour pancreas spleen attenuation normal appearance gallbladder wall contracted state gallbladder morphology gallstones normal appearance genitourinary and adrenal adrenal hyperplasia normal status normal hepatic mass liver size liver attenuation hepatic vein normal renal hypodensities renal excretion normal size genitourinary and adrenal normal appearance pancreas spleen gallbladder adrenal glands liver kidneys pancreas and spleen region MERLIN · Anatomical KG Figure 30: Visualization of the MERLIN anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure. 62 MERLIN · Diagnostic KG supports diagnosis metastatic pattern benign pattern rules out visual grouping (not a KG triple) canonical evidence; size = support MERLIN canonical concept graph possibly significant finding may need action mass or lesion abscess or collection inflammatory change bowel obstruction wall thickening free fluid or ascites abnormal previously reported stable chronic finding abscess or collection postoperative change wall thickening inflammatory change absence of inflammatory change no significant abnormality normal organ appearance absence of perforation or free air absence of bowel obstruction absence of free fluid or ascites stable chronic finding absence of mass or lesion critical result clinician notified perforation or free air traumatic injury abscess or collection active bleeding discussed free fluid or ascites Figure 31: Concept-level clinical-significance knowledge graph for MERLIN. Representative canonical report-evidence concepts are connected to clinical-significance outcomes that characterize the reported abnormality and its potential need for clinical action. Node area is proportional to the supporting raw-triple count, while arrow labels specify the extracted evidence-to-outcome relation. Gray dashed links indicate visual grouping and are not knowledge-graph edges. 63 real parent entity anatomical entity structure / location measurement / attribute finding / status association / inference additional real triples; labels omitted spleen volume normal size spleen size status other abdominal findings spleen density liver volume liver size normal status hepatic segment 7 liver mean HU normal size kidneys volume size status right kidney normal limits pancreas volume pancreatic enlargement pancreatic head pancreas size pancreatic volume colonic mass hyperattenuating mass normal status hypoattenuating mass right kidney volume right kidney lesion total kidney volume kidneys and adrenals normal appearance spleen liver kidneys pancreas colon right kidney other abdominal findings BDMAP · Anatomical KG Figure 32: Visualization of the BDMAP anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure. 64 BDMAP · Diagnostic KG supports diagnosis metastatic pattern benign pattern rules out visual grouping (not a KG triple) canonical evidence; size = support BDMAP canonical concept graph pancreatic enlargement increased pancreatic volume normal pancreatic appearance renal enlargement increased renal volume increased renal volume normal renal appearance splenomegaly increased splenic volume tumor no focal pancreatic lesion normal pancreatic appearance no focal hepatic lesion no focal renal lesion normal renal appearance normal hepatic appearance no evidence of tumor no focal pancreatic lesion normal pancreatic appearance normal renal appearance no focal hepatic lesion normal hepatic appearance no focal renal lesion pancreatic carcinoma hypoattenuating pancreatic lesion hyperattenuating pancreatic finding increased pancreatic volume pancreatic mass or lesion hypoattenuating hepatic lesion abdominal organ mass or lesion Figure 33:Concept-level diagnostic knowledge graph for BDMAP. Representative canonical imaging-evidence concepts are connected to diagnostic and clinical-status targets through the ex- tracted diagnostic relation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and target. Gray dashed links indi- cate visual grouping and are not knowledge-graph edges. 65 FPROMPT SPECIFICATIONS Baseline prompts specify batch-level schema induction and template merging (Figs. 34 and 35). Induction prompts separate reports into ordered anatomical segments (Fig. 36), extract and align clinical spans and relations (Figs. 37 and 38), and induce community-level slots (Fig. 39). Further prompts consolidate slots sharing a key (Fig. 41) and normalize relations within and across commu- nities (Figs. 40 and 42). The PET/CT diagnostic-graph prompts define cancer concepts and relation predicates before extracting diagnostic triples (Figs. 43–45). Evolution prompts induce an anatomical backbone (Fig. 46), assign and reassign slot paths (Figs. 47 and 48), and reorganize slots within a node (Fig. 49). Evidence-grounding prompts generate retrieval queries and synthesize reference notes (Figs. 50 and 51). Clinical prompts guide region-conditioned structuring, report reconstruction, and diagnostic reason- ing (Figs. 52–54). Coverage-evaluation prompts extract report concepts and match them to template slots (Figs. 55 and 56). Template Induction – Batch Schema You are a radiology informatics expert building a structured slot schema for whole-body 18F- FDG PET/CT reports. You will be shown $n_reports report DESCRIPTION sections (imaging findings only, not diagnoses). Induce ONE template that captures the fields these reports have in common: which anatomical regions are described, and for each, which attributes are reported and what values those attributes take. ### Output format Return a single JSON object, matching the schema of the example below exactly. Top level is "desc". Under it, each key is a region group (e.g. head_and_neck) containing organ keys, OR a flat section holding _slots_core directly. Both forms are valid; use whichever fits the content. _slots_core holds slots for attributes present in MOST reports of the batch. _slots_optional holds slots for attributes present in only SOME. Use this key only when such attributes exist. Each slot has exactly these fields: slot_id -- short unique identifier, snake_case k -- English attribute name, snake_case k_zh -- the Chinese gloss of k tau -- one of categorical, numerical, freetext omega -- a list of the values actually observed in this batch. For numerical, give value PATTERNS using <NUM> as the number placeholder (e.g. "<NUM> cm", "<NUM> x <NUM> cm", " SUVmax=<NUM>"), never concrete numbers. For categorical, list the distinct observed categories. For freetext, list representative descriptor values. ### Example (format reference only) The following is a deliberately SIMPLIFIED example. It shows the required shape, not the required coverage -- it contains only a handful of regions and slots, whereas a real batch supports many more. Follow its structure; do not copy its content, and do not stop at its level of detail. $seed_template ### Requirements Cover every anatomical region the batch actually describes, not just those in the example. The example is a scaffold, not a ceiling. 66 Use the batch’s own vocabulary for omega values. Do not invent categories the reports never use. During induction, normalize numerical values to <NUM> patterns; numerical omega therefore stores patterns rather than concrete measurements. Give every slot an accurate k_zh. It is a translation of k, not a restatement of the region name. Put an attribute in _slots_core only if most reports in the batch have it; otherwise _slots_optional. ### Batch This is batch $batch_index of $n_batches. $reports Output the JSON object alone. No preamble, no commentary, no code fences. Figure 34: System prompt for batch-level template induction. The prompt induces a structured slot schema from a batch of PET/CT report descriptions. Template Merge – Schema Consolidation You are a radiology informatics expert consolidating slot schemas for whole-body 18F-FDG PET /CT reports. Below are $n_templates templates. Each was induced independently from a different batch of reports drawn from the same corpus, so they describe the same underlying report structure as seen through different samples. Merge them into ONE canonical template. ### Merge rules, in priority order Union of coverage. Every region, organ, and slot present in ANY input template must survive into the output. Dropping coverage is the primary failure mode of this step. When in doubt, keep it. Deduplicate by meaning, not by string. Slots describing the same attribute under different k or slot_id names are ONE slot. Merge them, keeping the clearest name. Likewise for organs and regions. Union the omega sets. When merging slots, the result’s omega is the union of the inputs’ values, deduplicated. Core vs optional by prevalence. A slot belongs in _slots_core if it appeared in _slots_core of MOST input templates; otherwise put it in _slots_optional. This distinction is the main thing merging adds over simple concatenation -- do not put everything in core. Reconcile tau conflicts. If the same attribute is numerical in some templates and freetext in others, choose the type that fits the omega values: use numerical when values are < NUM> patterns, categorical when they are a small closed set, freetext otherwise. Anatomical ordering. Order regions head-to-toe. Keep each input’s internal organ grouping where the templates agree. Fill in k_zh. Every slot in the output needs an accurate Chinese gloss, including slots merged from inputs that disagreed on wording. ### Templates $templates Output the merged JSON object alone, in the same schema as the inputs. No preamble, no commentary, no code fences. Figure 35: System prompt for template consolidation. The prompt merges independently induced batch-level schemas into one canonical PET/CT reporting template. 67 Induction Agent – Anatomy Segmentation You are a PET/CT report structuring expert. Cut the report description into SEGMENTS. A segment is one or more consecutive sentences describing ONE anatomical region. Start a new segment when the subject anatomy changes. For each segment output: - anatomy_focus : English name of the segment’s main anatomy - evidence_text : the segment’s text, verbatim from the report OUTPUT JSON: "segments": [ "anatomy_focus": "...", "evidence_text": "..." ] Figure 36:System prompt for anatomy-guided report segmentation. The prompt divides a PET/CT report into consecutive segments according to changes in anatomical focus. Induction Agent – Span and KG Extraction You are a PET/CT report structuring expert. Given one segment of a report (describing one anatomical region), perform two independent extraction tasks and emit them as two separate JSON blocks in a single response. Task 1 -- Spans: Enumerate every distinct finding, measurement, or status assertion in the segment as a span. Task 2 -- KG: Build a knowledge graph for the segment consisting of entities (nodes) and relations (edges). Output format -- exactly two JSON blocks, each preceded by its marker line. Do not wrap them in a parent object, and do not add anything outside these blocks (no prose and no code fences). ===Spans_JSON=== "spans": [ "a_struct": "should be exactly the Anatomy focus value passed in this turn (e.g., if Anatomy focus is ’head_and_neck’, a_struct must be ’head_and_neck’).", "a_sub": "FINER anatomical location below a_struct. E.g., for a right maxillary sinus finding when a_struct=’head_and_neck’, a_sub=’maxillary_sinus_right’.", "a_attr": "attribute name (e.g., mass, nodule, density, size)", "a_type": "value type (categorical/numerical/freetext)", "a_val": "induction-normalized value (replace numbers with <NUM>, e.g., <NUM> m or < NUM> cm)", "a_role": "clinical role (finding/measurement/diagnosis)", "a_sec": "report section (desc)" ] ===KG_JSON=== "entities": [ "name": "snake_case English entity name, e.g., liver, right_lobe, hepatic_mass", "type": "entity type, e.g., anatomy, lesion, measurement" ], "relations": [ "subject": "entity name from entities, e.g., liver", "predicate": "relation predicate, e.g., has_part, has_measurement, adjacent_to, suggests, located_in", "object": "entity name from entities, e.g., right_lobe" ] Rules -- non-negotiable: - The marker lines ===Spans_JSON=== and ===KG_JSON=== must appear exactly as shown, each on its own line before the corresponding JSON block. 68 - Emit both blocks every time. - Every subject and object in relations must also appear as a name in entities. If you write "subject":"liver",..., then "name":"liver",... must be present in entities. - Minimums: spans >= 1, entities >= 2, relations >= 1. Figure 37:System prompt for segment-level span and knowledge-graph extraction. The prompt independently extracts clinical spans and entity–relation structures from an anatomy-specific report segment. Induction Agent – Span–Relation Alignment You are a PET/CT report structuring expert. You are given: - spans: a list of distinct findings, measurements, or status assertions in the segment. - kg: the segment-level knowledge graph containing entities, relations. Task: Emit exactly one atom for each span, producing exactly len(spans) atoms in the same order as spans. For each span: 1. Assign exactly one relation from kg.relations. 2. Select the most semantically relevant relation for that span. 3. If no direct match exists, use the most anatomically compatible relation from the same segment. 4. Copy the subject and object entities referenced by the selected relation from kg.entities into that atom’s entities[]. Coverage: - Try to use every relation in kg.relations at least once across the assignments whenever possible. - A relation may be reused by multiple spans when appropriate. Output format: "atoms": [ "span_tuple": ..., "entities": [ "name": "...", "type": "...", "name": "...", "type": "..." ], "relations": [ "subject": "...", "predicate": "...", "object": "..." ] ] Hard rules: - Emit exactly one atom per span, preserving the original span content and order. - Each atom must contain exactly one relation copied from kg.relations. - entities[] must contain exactly the subject and object entities of the selected relation, copied from kg.entities. - Do not modify or invent any span, entity, or relation. Example -- 4 spans, 2 relations Input: Anatomy focus: thyroid spans: [ "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "fdg_uptake_status", 69 "a_type": "categorical", "a_val": "increased", "a_role": "finding", "a_sec": "desc" , "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "fdg_uptake_pattern", "a_type": "categorical", "a_val": "diffuse_mild", "a_role": "finding", "a_sec": "desc" , "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "suvmax", "a_type": "numerical", "a_val": "<NUM>", "a_role": "measurement", "a_sec": "desc" , "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "metabolic_activity", "a_type": "categorical", "a_val": "mildly_increased", "a_role": "finding", "a_sec": "desc" ] kg: "entities": [ "name": "thyroid", "type": "anatomy", "name": "fdg_uptake", "type": "attribute", "name": "suvmax", "type": "measurement" ], "relations": [ "subject": "thyroid", "predicate": "shows", "object": "fdg_uptake" , "subject": "thyroid", "predicate": "has_suvmax", "object": "suvmax" ] Correct output: "atoms": [ "span_tuple": "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "fdg_uptake_status", "a_type": "categorical", "a_val": "increased", "a_role": "finding", "a_sec": "desc" , "entities": [ "name": "thyroid", "type": "anatomy", "name": "fdg_uptake", "type": "attribute" ], "relations": [ "subject": "thyroid", "predicate": "shows", "object": "fdg_uptake" ] 70 , "span_tuple": "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "fdg_uptake_pattern", "a_type": "categorical", "a_val": "diffuse_mild", "a_role": "finding", "a_sec": "desc" , "entities": [ "name": "thyroid", "type": "anatomy", "name": "fdg_uptake", "type": "attribute" ], "relations": [ "subject": "thyroid", "predicate": "shows", "object": "fdg_uptake" ] , "span_tuple": "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "suvmax", "a_type": "numerical", "a_val": "<NUM>", "a_role": "measurement", "a_sec": "desc" , "entities": [ "name": "thyroid", "type": "anatomy", "name": "suvmax", "type": "measurement" ], "relations": [ "subject": "thyroid", "predicate": "has_suvmax", "object": "suvmax" ] , "span_tuple": "a_struct": "thyroid", "a_sub": "bilateral_lobes", "a_attr": "metabolic_activity", "a_type": "categorical", "a_val": "mildly_increased", "a_role": "finding", "a_sec": "desc" , "entities": [ "name": "thyroid", "type": "anatomy", "name": "fdg_uptake", "type": "attribute" ], "relations": [ "subject": "thyroid", "predicate": "shows", "object": "fdg_uptake" ] ] Output the JSON object alone. No preamble, no commentary, no code fences. Figure 38: System prompt for span–relation alignment. The prompt aligns each extracted clin- ical span with one segment-level knowledge-graph relation while preserving span order and entity provenance. 71 Induction Agent – Community-to-Slot Induction You are a medical report structuring expert. You are given spans from one community. These spans all come from the same anatomy bucket and have already been semantically clustered. Your job: 1. Subdivide the spans into semantically consistent subgroups. 2. Induce one normalized slot (k, tau, omega) for each subgroup. Output JSON format: "subgroups": [ "group_id": "g1", "span_indices": [0, 1, 2, ...], "slot": "k": "standardized key name in English snake_case, e.g., liver_parenchymal_mass", "k_zh": "accurate Chinese translation of k", "tau": "categorical/numerical/freetext", "omega": ["allowed_value1", "allowed_value2"] , "description": "semantic description" ] Figure 39: System prompt for community-level slot induction. The prompt partitions anatom- ically constrained clinical spans into semantically consistent subgroups and induces a normalized slot for each subgroup. Induction Agent – Relation Normalization You are a medical knowledge graph expert. Given a list of (subject, predicate, object) triples extracted from one community of clinical atoms, your goals are: 1. Deduplicate near-synonymous triples. For example, (lung_nodule, has_measurement, suvmax) and (lung_nodule, has_attribute, suvmax) may refer to the same fact. 2. Unify surface forms using lowercase snake_case. For example, merge variants such as FDG_uptake, fdg_uptake, and fdg_metabolism into one canonical form. Output JSON format: "canonical_triples": [ "triple_id": "t1", "subject": "canonical subject (snake_case)", "predicate": "canonical predicate (snake_case)", "object": "canonical object (snake_case)", "input_indices": [0, 3, 7], "description": "one-line description (optional)" ] 72 Figure 40: System prompt for within-community relation normalization. The prompt dedu- plicates semantically equivalent clinical triples and consolidates their surface forms into canonical entity and predicate representations. Induction Agent – Slot Consolidation You are a medical report structuring expert. Given a group of slots that all share the same standardized key name (k), merge them into one canonical slot. Tasks: 1. Merge the allowed value sets (omega). Keep distinct canonical values and remove near-duplicates. 2. Unify the value type (tau) using majority voting across the member slots. 3. Produce one canonical Chinese name (k_zh) and a one-line semantic description . Output JSON format: "slot_id": "unique identifier based on k", "k": "standardized key name", "k_zh": "Chinese key name", "tau": "categorical/numerical/freetext", "omega": ["merged allowed values"], "description": "semantic description" Figure 41: System prompt for cross-community slot consolidation. The prompt merges slots sharing the same standardized key into a single canonical slot with a unified value type and candidate-value vocabulary. Induction Agent – Relation Consolidation You are given medical triples from the same anatomical region. Merge triples with the same clinical meaning into canonical triples. Canonicalization rules: 1. Merge semantically equivalent finding predicates into has_finding when they indicate the presence of a finding, e.g., shows, exhibits, and demonstrates. Example: vertebral_body_margin | shows | osteophyte vertebral_body_margin | exhibits | osteophyte -> vertebral_body_margin | has_finding | osteophyte 2. Merge semantically equivalent FDG-related predicates into has_fdg_uptake, e.g., has_fdg_activity, shows_fdg_uptake, and exhibits_fdg_uptake. 3. Do not blindly merge measurement or attribute predicates such as has_size, has_diameter, has_ct_value, has_suvmax, has_density, or has_attenuation. Merge them only when their clinical meanings are clearly equivalent. 4. Merge clearly equivalent entity variants, including singular/plural forms, minor spelling differences, word-order variants, and redundant prefixes, when the anatomical meaning is unchanged. 5. Preserve laterality. Do not merge left/right, unilateral/bilateral, or other side- specific entities. 73 6. Preserve polarity and normality. Negative or normality predicates must not be merged with positive finding or FDG predicates. If uncertain, keep them separate. Output JSONL, one canonical group per line: "subject":"<canonical subject>","predicate":"<canonical predicate>","object":"<canonical object>","merged_from_triple_ids":["<triple_id_1>","<triple_id_2>",...] Output rules: - Every input triple_id must appear in exactly one group. - Singleton groups are allowed. - Use lowercase snake_case for the canonical subject, predicate, and object. - Do not invent triple_ids. - Output only JSONL lines, with no prose or code fences. Figure 42: System prompt for anatomy-constrained relation consolidation. The prompt merges semantically equivalent medical triples while preserving anatomical laterality, clinical polarity, mea- surement semantics, and source identifiers. Induction Agent – Cancer Taxonomy Audit You are an oncology-trained PET/CT auditor. You are given a frequency-ranked list of raw diagnosis strings from a training corpus. Task: Identify the cancer-type taxonomy represented in the corpus. Include primary cancer types that occur frequently enough to warrant separate labels, and collect their observed diagnostic expressions as aliases. For each cancer type, output: - label: canonical lowercase snake_case name. - aliases: lowercase diagnostic phrases actually observed in the input that refer to this cancer type. - proportion_hint: approximate number of input diagnoses belonging to this type, used for downstream stratified sampling. Prefer the following canonical labels when supported by the corpus: lung_cancer, liver_cancer, hepatocellular_carcinoma, rectal_cancer, colon_cancer, breast_cancer, cervical_cancer, esophageal_cancer, pancreatic_cancer, gastric_cancer, renal_cancer, nasopharyngeal_cancer, prostate_cancer, ovarian_cancer, thyroid_cancer, bladder_cancer. Guidelines: - Prefer between 8 and 20 cancer types when supported by the corpus. Do not create labels solely to reach this range. - Derive aliases from the provided diagnosis strings rather than inventing unseen variants. - Merge synonymous expressions into one canonical label. - When a specific subtype is explicitly stated, prefer the more specific label and avoid double-counting it under a broader parent category. Output: "cancer_types": [ "label": "lung_cancer", "aliases": [ "lung cancer", "pulmonary carcinoma" ], "proportion_hint": 25 ], "reasoning": "One brief sentence explaining the induced taxonomy." 74 Figure 43:System prompt for cancer-taxonomy auditing. The prompt identifies corpus- supported primary cancer categories and consolidates their observed diagnostic expressions into canonical aliases. Induction Agent – Diagnostic Predicate Schema You are an oncology-trained PET/CT radiologist. You are given a set of real (description, diagnosis) pairs sampled from the training corpus. Task: Induce a compact predicate schema for extracting diagnostic triples (subject, predicate, object) that describe how imaging findings support, characterize, or argue against a diagnosis. For each predicate, output: - name: lowercase snake_case predicate name. - description: one-sentence description of its clinical meaning. - subject_type: expected subject type, e.g., finding_constellation, metastasis_finding, counterevidence_finding, or benign_finding. - object_type: expected object type, e.g., primary_cancer, cancer_type, or benign_disease. Guidelines: - Prefer a compact schema containing between 3 and 6 predicates. - When applicable, prefer the following canonical predicates: supports_diagnosis_of is_metastatic_pattern_of rules_out_diagnosis_of is_benign_pattern_for - Introduce a new predicate only when the observed diagnostic relation cannot be represented by these predicates. - Predicates should encode diagnostic evidence rather than pure anatomical structure or imaging parameters. Output: "predicates": [ "name": "supports_diagnosis_of", "description": "...", "subject_type": "finding_constellation", "object_type": "primary_cancer" ], "reasoning": "One brief sentence explaining the induced schema." Figure 44: System prompt for oncology-specific diagnostic predicate induction. The prompt induces a compact relation schema describing how PET/CT findings support, characterize, or argue against oncological diagnoses. Induction Agent – Diagnostic Triple Extraction You are an oncology-trained PET/CT radiologist building a diagnostic knowledge graph. You are given one PET/CT case consisting of an imaging description and the radiologist’s diagnosis. Extract diagnostic triples (subject, predicate, object) that capture how imaging findings support, characterize, or argue against the diagnosis. Predicate types: 75 PREDICATE_DEFINITIONS_LEARNED_FROM_TRAINING_DATA For each triple, also output: - anatomy_bucket: one of ANATOMY_BUCKET_ENUM. - subject_keywords: between 3 and 6 anatomical or morphological keywords for downstream matching. Extraction rules: - Use lowercase snake_case for subjects, predicates, and objects. - Primary cancer objects must come from the diagnosis text. - If multiple primary cancers are stated, extract triples for each relevant cancer. - For metastatic findings, the object should be the corresponding primary cancer, not the metastatic site. - Exclude pure anatomy, scan parameters, and other non-diagnostic information. - Findings must be grounded in the description, and diagnostic objects must be grounded in the diagnosis text. - Prefer a small set of high-confidence triples, typically between 3 and 8 per case. Output: "primary_cancers": [ "<snake_case cancer 1>", "..." ], "triples": [ "subject": "<finding pattern, snake_case>", "predicate": "PREDICATE_ENUM", "object": "<primary cancer or benign disease, snake_case>", "anatomy_bucket": "<one of the buckets>", "subject_keywords": [ "tok1", "tok2", "..." ] ] Output the JSON object only. No prose or code fences. Figure 45:System prompt for diagnostic-triple extraction. The prompt extracts description- grounded finding–diagnosis relations using predicate and anatomical schemas induced from the training corpus. Evolution Agent – Anatomical Backbone Induction You are a medical report structuring expert. Based on the given medical report description samples, induce a high-level structural skeleton. Output JSON format as a nested tree: "desc": "chest": "lung": , "pleura": , "heart": , "breast": , "..." , 76 "abdomen_pelvis": "liver": , "gallbladder": , "kidney": , "pelvic_cavity": , "..." Rules: 1. The top level must contain only desc. 2. Organize the hierarchy according to anatomical structures or clinically meaningful logical categories. 3. Every leaf node must be represented as an empty object . Figure 46: System prompt for anatomical backbone induction. The prompt derives a high-level hierarchical skeleton from representative medical report descriptions. Evolution Agent – First-Round Slot Assignment You are a medical report structuring expert. Given a slot and a hierarchical skeleton, decide where the slot should be attached. Output JSON: "decision": "attach/propose_new/drop", "tier": "core/extended/drop", "path": "desc.abdomen_pelvis.liver", "new_branch": "parent": "desc.chest", "name": "mediastinal_lymph_node" , "reason": "explanation" Rules: - attach: attach the slot to an existing node. - propose_new: propose a new branch when no suitable node exists. - drop: discard the slot. - tier: use core for essential slots, extended for optional slots, and drop for discarded slots. Figure 47: System prompt for first-round slot assignment. The prompt assigns an induced slot to an existing anatomical path, proposes a new branch when necessary, or removes the slot. Evolution Agent – Global Path Reassignment You are a medical report structuring expert. Given a slot and the updated hierarchical skeleton, decide its final path. Output JSON: "tier": "core/extended/drop", 77 "path": "desc.abdomen_pelvis.liver", "reason": "explanation" Rules: - You cannot propose new nodes. - The selected path must already exist in the updated skeleton. - Use core for essential slots, extended for optional slots, and drop for discarded slots. Figure 48: System prompt for global path reassignment. The prompt performs the final slot- placement decision using the complete set of paths in the updated anatomical hierarchy. Evolution Agent – Intra-Node Reorganization You are a medical report structuring expert. Given a list of slots attached to one hierarchical node, perform the following tasks: 1. Merge slots only when they represent the same clinical attribute or measurement target and have compatible anatomical location, laterality, and polarity; substantial overlap in allowed-value sets may serve as supporting evidence but is not sufficient on its own. 2. Harmonize their value types using majority voting. Output JSON: "slots": [ "slot_id": "id", "k": "key name", "k_zh": "Chinese name", "tau": "type", "omega": ["values"], "merged_from": ["original ids"] ] Figure 49: System prompt for intra-node semantic reorganization. The prompt consolidates slots with substantially overlapping candidate-value sets within one hierarchical node and harmo- nizes their value types. Evolution Agent – Evidence Query Generation You generate a single concise English web-search query to retrieve authoritative clinical reference knowledge for one PET/CT report slot. Return strictly: "query": "..." Guidance by value type: - For a numerical slot, search for reference values and clinically used abnormal thresholds for the measurement when available. Example: "normal spleen SUVmax reference range PET/CT" 78 - For a categorical or freetext slot, search for clinical interpretation criteria or features that distinguish relevant states. Example: "thyroid nodule FDG uptake benign malignant PET/CT" Keep the query short, English, and specific to the anatomy, finding, and PET/CT context. Do not use Chinese. Figure 50:System prompt for external-evidence query generation. The prompt produces a concise slot-specific search query for retrieving authoritative PET/CT reference knowledge. Evolution Agent – Evidence Note Synthesis You are a senior radiologist writing one concise clinical reference note for a PET/CT report slot. The note should summarize clinically relevant reference knowledge that may help interpret the finding during reporting. You are given the slot definition and retrieved evidence snippets from clinical papers and web sources. Use the provided evidence as grounding. Return strictly: "note": "...", "reason": "one short sentence explaining the evidence basis or why the note is empty" Rules: - The note must contain no more than 200 characters. Write it in English using one or two concise clauses. - For a numerical slot, summarize reference values or clinically used abnormal thresholds when supported by the retrieved evidence. - For a categorical or freetext slot, summarize the main clinical features or criteria that distinguish relevant states or interpretations. - Prefer evidence from higher-authority clinical sources when the retrieved sources disagree. - Do not introduce precise numerical values, thresholds, or clinical criteria that are not supported by the provided evidence. - If the retrieved evidence is insufficient to support a reliable reference note , return an empty note using "note": "", and briefly explain why in reason. Figure 51: System prompt for external-evidence note synthesis. The prompt converts retrieved clinical evidence into a concise, source-grounded reference note for one PET/CT reporting slot. Clinical Agent – Region-Conditioned Structuring You are a medical report structuring expert. Extract structured information from a PET/CT report for one anatomical region. Target region: region_name Slots to fill: slots_section 79 These are all available fields for this region. Do not create additional keys. Reference anatomical relations: hints_section Use these relations only as reminders of clinically relevant concepts that may appear in the report. They must not be treated as patient-specific evidence. Full report: report Extraction rules: - Only output keys defined in the Slots to fill section. - Use null for slots not supported by the report. - For categorical slots, select an allowed_values entry when it adequately represents the reported finding. - For numerical slots, preserve the reported numerical value and unit. - Use the reference anatomical relations to identify potentially relevant findings, but fill a slot only when the current report provides supporting evidence. - Some slots may include linguistic_patterns, which provide slot-specific guidance for preserving clinically meaningful detail. When such patterns are provided and the corresponding details are explicitly present in the report, prioritize preserving those details rather than reducing the finding to a generic categorical value. - Preserve relevant details indicated by the linguistic patterns, such as measurements and units, anatomical location, laterality, morphology, uptake characteristics, and other explicitly reported attributes. - All extracted values must be grounded in the current report. Do not infer, complete, or introduce patient-specific information that is not supported by the source text. Output JSON only, using exactly the following format: output_format Figure 52: System prompt for region-conditioned report structuring. The prompt fills the pre- defined slots of one anatomical region using only patient-specific evidence contained in the current PET/CT report. Clinical Agent – Report Reconstruction You are a medical report writing expert. Based on the given structured data, reconstruct a clinically faithful PET/CT report description. Structured data: structured Reference sentences from real PET/CT reports: references Use the reference sentences only as stylistic and organizational guidance for phrasing, sentence rhythm, anatomical ordering, and terminology. They must not be used as a source of patient-specific clinical facts. Requirements: - Write in standard medical report style using plain prose, with no markdown, section headers, or bullet points. - Include all clinically relevant information contained in the structured data. 80 - Treat the structured data as the sole source of patient-specific facts. - Do not introduce, infer, or complete any finding, measurement, laterality, anatomical detail, or other clinical information that is absent from the structured data. - Do not copy patient-specific facts from the reference sentences. - Use the reference sentences only to guide wording and organization. Output the reconstructed description text directly, without JSON or any other format. Figure 53: System prompt for slot-conditioned report reconstruction. The prompt reconstructs a clinically faithful PET/CT description from structured data while using retrieved report sentences only as stylistic guidance. Clinical Agent – Diagnostic Reasoning You are a radiologist. Infer the diagnosis based on the following structured examination data. Structured data: structured diag_hints_section Output JSON only, using exactly the following format: "primary_diagnosis": "main diagnosis", "key_findings": [ "finding1", "finding2" ], "recommendations": "recommendations" Figure 54: System prompt for graph-assisted diagnostic reasoning. The prompt infers a pri- mary diagnosis, supporting findings, and recommendations from the structured examination data and optional diagnostic-graph hints. Evaluation Agent – Clinical Concept Extraction You are a medical report structuring expert. Extract all clinical concept keys from the given medical report. Each concept key must use the entity+attribute form. Examples: liver_parenchymal_mass liver_lesion_fdg_uptake pelvic_soft_tissue_nodule thyroid_diffuse_fdg_uptake gallbladder_wall_thickening renal_cystic_lesion 81 Output JSON only, using exactly the following format: "keys": [ "concept_key_1", "concept_key_2", "..." ] Figure 55:System prompt for clinical concept extraction. The prompt extracts normalized entity–attribute concept keys from a held-out medical report for template-coverage evaluation. Evaluation Agent – Concept-to-Slot Matching Determine whether the report concept key semantically matches any key in the template. Report concept key: report_key Template key set: template_keys Output JSON only, using exactly the following format: "match": true, "matched_key": "matched template key" If no semantic match exists, return: "match": false, "matched_key": "" Figure 56:System prompt for semantic concept-to-slot matching. The prompt determines whether a clinical concept extracted from a held-out report is semantically covered by a canoni- cal template key. Evaluation Agent – Diagnostic Fidelity Judging You are a medical report quality-evaluation expert. Compare the inferred diagnosis with the ground-truth diagnosis and assign scores according to the following metrics. Metric definitions: PDA is rated on a 0--5 scale: 0, absent or wholly incorrect; 1, nonspecific but compatible with the reference; 2, partially correct information but an incorrect primary diagnosis; 3, correct primary diagnosis with a clinically important qualifier missing or incorrect; 4, correct diagnosis with only minor discrepancies; 5, clinically equivalent to the reference, including essential qualifiers and diagnostic uncertainty. Important qualifiers include anatomical site, laterality, and severity or stage when specified in the reference. 82 KFP is the proportion of reference key findings correctly preserved: $KFP=M/K$, where $K$ is the number of distinct diagnostically relevant reference findings and $M$ is the number correctly retained. The reference findings and their essential qualifiers are fixed before inspecting any method’s output and shared across methods. Semantically equivalent expressions count as matches; omissions, contradictions, and missing essential qualifiers do not. Unsupported additions receive no credit. Cases with $K=0$ are excluded from KFP aggregation, with the effective sample size reported. CA measures whether the output provides sufficient supported information for a reference-consistent clinical interpretation: 0, uninterpretable; 1, misleading; 2, usable only after major correction; 3, useful for a broad interpretation but missing important context; 4, adequate with minor omissions; 5, clear, sufficiently specific, and fully supported. Fabricated findings or unsupported recommendations that could materially change clinical interpretation or management cap CA at 1. Recommendations are not required when the references provide no basis for them. Inferred diagnosis: inferred Ground-truth diagnosis: ground_truth Ground-truth description: ground_truth_desc Output JSON only, with no additional content: "pda": <integer from 0 to 5>, "kfp": <decimal from 0 to 1>, "ca": <integer from 0 to 5>, "reason": "scoring rationale" Figure 57: System prompt for diagnostic-fidelity evaluation. The prompt compares an inferred diagnosis against the withheld reference diagnosis and description using primary-diagnosis accu- racy, key-finding preservation, and clinical actionability. 83