Paper deep dive
EviDent-CBCT: Evidence-Bottlenecked Report Generation from Dental CBCT under Non-Exhaustive Report Supervision
Ruiyang Hao, Zhi Qin Tan, Yulan He, Owen Addison, Yunpeng Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/5/2026, 5:50:53 AM
Summary
The paper introduces EviDent-CBCT, an evidence-bottlenecked framework for generating dental CBCT reports under non-exhaustive supervision. It utilizes an anatomy-aware network to predict discrete tooth-level, global, and tooth-IAC evidence, applies dental-logic consistency projection, and employs an image-blind local language model for report generation. The method achieves superior performance compared to direct baselines and ranked highly in the ODIN 2026 challenge.
Entities (8)
Relation Signals (5)
EviDent-CBCT → achievedrankin → ODIN 2026
confidence 90% · In the ODIN 2026 challenge, it ranked second in automated evaluation and third in blinded clinical Arena comparison
EviDent-CBCT → evaluates → RadFact-Lite-Dental
confidence 90% · achieves 0.402 ± 0.003 RadFact-Lite-Dental logical-F1
EviDent-CBCT → trainedon → ToothFairy4
confidence 90% · The paired subset of ToothFairy4 contains 624 CBCT volumes... The method isolates these measurable interfaces... Stage 2 maps anatomy-aligned inputs
EviDent-CBCT → uses → Qwen2.5-7B
confidence 90% · We distill the teacher’s evidence-to-report realization behavior into a local Qwen2.5-7B rewriter
EviDent-CBCT → predicts → FDI tooth position
confidence 85% · E y tooth contains findings indexed by FDI tooth position
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Dento-maxillofacial cone-beam CT (CBCT) reports may contain dozens of tooth-specific, anatomical, and spatial findings from a single 3D scan. Learning to generate such reports from limited clinical data is challenging because routine reports may not exhaustively document image findings, and a non-mention may reflect either absence or non-reporting. We present EviDent-CBCT, an evidence-bottlenecked framework designed for this incomplete supervision. An anatomy-aware network maps each CBCT scan to a discrete record of tooth-level, global, and tooth-IAC evidence. A dental-logic consistency projection reconciles incompatible evidence before a deterministic renderer and an image-blind local language model generate the report using only this record. For tooth-level evidence, reliability-aware training uses eligible non-mentions as reduced-weight negatives, while unreported global and tooth-IAC labels remain unknown. A metal-sensitive input channel preserves intensity cues from dental materials. Across three validation runs, EviDent-CBCT achieves $0.666\pm0.006$ merged evidence set-F1 and $0.402\pm0.003$ RadFact-Lite-Dental logical-F1, versus $0.371\pm0.018$ for the strongest controlled direct baseline. In the ODIN 2026 challenge, it ranked second in automated evaluation and third in blinded clinical Arena comparison on the hidden test set. These results support the discrete evidence record as an effective and auditable interface for CBCT report generation.
Tags
Links
- Source: https://arxiv.org/abs/2610.02375v1
- Canonical: https://arxiv.org/abs/2610.02375v1
Trouble viewing inline? Open PDF directly →
Full Text
102,887 characters extracted from source content.
Expand or collapse full text
EviDent-CBCT: Evidence-Bottlenecked Report Generation from Dental CBCT under Non-Exhaustive Report Supervision Ruiyang Hao, Zhi Qin Tan, Yulan He, Owen Addison, Yunpeng Li ∗ King’s College London ruiyang.hao, zhi_qin.tan, yulan.he, owen.addison, yunpeng.li@kcl.ac.uk Abstract Dento-maxillofacial cone-beam CT (CBCT) reports may con- tain dozens of tooth-specific, anatomical, and spatial findings from a single 3D scan. Learning to generate such reports from limited clinical data is challenging because routine re- ports may not exhaustively document image findings, and a non-mention may reflect either absence or non-reporting. We present EviDent-CBCT, an evidence-bottlenecked framework designed for this incomplete supervision. An anatomy-aware network maps each CBCT scan to a discrete record of tooth- level, global, and tooth-IAC evidence. A dental-logic consis- tency projection reconciles incompatible evidence before a de- terministic renderer and an image-blind local language model generate the report using only this record. For tooth-level evi- dence, reliability-aware training uses eligible non-mentions as reduced-weight negatives, while unreported global and tooth- IAC labels remain unknown. A metal-sensitive input chan- nel preserves intensity cues from dental materials. Across three validation runs, EviDent-CBCT achieves 0.666± 0.006 merged evidence set-F1 and 0.402 ± 0.003 RadFact-Lite- Dental logical-F1, versus 0.371± 0.018 for the strongest con- trolled direct baseline. In the ODIN 2026 challenge, it ranked second in automated evaluation and third in blinded clinical Arena comparison on the hidden test set. These results sup- port the discrete evidence record as an effective and auditable interface for CBCT report generation. 1 Introduction Cone-beam computed tomography (CBCT) provides three- dimensional views of dento-maxillofacial structures and is widely used for dental diagnosis and treatment plan- ning (Kaasalainen et al. 2021; Reddy et al. 2024). A single scan may reveal findings involving individual teeth, the jaws, sinuses, scan coverage, and spatial relations to the inferior alveolar canal (IAC), often requiring assignment to a specific anatomical region or FDI tooth position. CBCT reporting is therefore more than image captioning: a system must recover case-specific findings, preserve their anatomical identity and polarity, and express them coherently. Routine reports, however, record what was documented in clinical practice rather than exhaustively annotating the image. A non-mention may indicate true absence, report- specific omission, habitual non-reporting, or oversight, and these mechanisms cannot be distinguished from the report ∗ Corresponding author. This work was supported by King’s College London via the Faculty of Dentistry, Oral and Craniofacial Sciences Newland Pedley Studentship. alone (Pandharipande et al. 2016; Hartung et al. 2020). Nev- ertheless, aggregate non-mention patterns may provide use- ful negative signal without establishing absence in an indi- vidual case. We refer to this setting as non-exhaustive routine report supervision; its scope and interpretation are detailed in supplementary Sect. A. This setting is particularly challenging for direct image- to-text generation. Recent radiology systems connect visual encoders to large language models (Wang et al. 2023; Li et al. 2024; Liu et al. 2024a; Xiao et al. 2025), but with limited CBCT-report pairs, a direct generator must jointly learn three-dimensional anatomy, visual grounding, report- specific content selection, and language generation from in- complete targets. These functions remain entangled, obscur- ing whether candidate-reference discrepancies arise from vi- sual prediction, incomplete documentation, or realization. Canonicalization reduces linguistic variation (Vosshenrich et al. 2023; Bannur et al. 2023) but not sparse visual supervi- sion; auxiliary evidence losses do not ensure that generation depends on predicted evidence; and fixed templates provide control at the cost of expressive flexibility (Delbrouck et al. 2025; Li et al. 2025). These limitations motivate an explicit content interface that separates evidence prediction from lan- guage realization and prevents the decoder from bypassing predicted findings. We propose EviDent-CBCT, a three-stage framework built around a discrete evidence bottleneck (Figure 1). Stage 1 segments individual teeth and key anatomical structures and extracts structure-specific features. Stage 2 uses anatomy- aligned branches to predict tooth-level, global, and tooth-IAC evidence; eligible tooth-level non-mentions provide reduced- weight negative supervision, unreported global and tooth- IAC labels remain unknown, and a metal-sensitive chan- nel preserves dental-material intensity cues. A determinis- tic dental-logic projection then reconciles incompatible tu- ples using anatomical semantics and confidence. Stage 3 renders the corrected record into a deterministic draft and supplies only this draft to an image-blind language model. Thus, unlike direct generators conditioned on continuous vi- sual representations, EviDent-CBCT makes the consistency- projected record the sole source of case-specific informa- tion, preventing bypass while keeping content selection and rewrite-induced deviations auditable. Across three validation runs, EviDent-CBCT achieves 0.402± 0.003 RadFact-Lite- arXiv:2610.02375v1 [cs.CV] 1 Oct 2026 DENTAL CBCT INPUT Dental CBCT Axial Coronal Sagittal FDI Tooth Indexing EVIDENCE SUPERVISION MODE Tooth branch supervision With reliability-aware non-mention supervision Global branch supervision Only explicit supervision Relation branch supervision Only explicit supervision Accounting for Routine Clinical Reporting Practice STAGE 3: EVIDENCE-BOTTLENECKED REPORT GENERATION Deterministic Renderer (from evidence tuples) Canonical Evidence Draft (structured, deterministic) Image-blind Rewrite (local LLM - Qwen) Final Clinical Report GPT Teacher Rewriter (cloud LLM) Qwen Student Rewriter (local LLM) rewrite distillation (training only) Inference Path Training Supervision Unmentioned Supervision STAGE 1: ANATOMICAL COORDINATE SYSTEM Where is each dental and maxillofacial structure? Input 3D UNet-like U-Mamba2 Encoder-decoder Output Masks (46 anatomical structures) Structure Tokens (Compact Representations) Tooth Mandible Maxilla Sinus IAC Others CONSISTENCY-PROJECTED EVIDENCE BOTTLENECK STAGE 2: Evidence Prediction What is the symptom or condition in dental and maxillofacial structure? TOOTH BRANCH 32FDIpositions 11 findings GLOBAL BRANCH Whole-Volume evidence 12 findings RELATION BRANCH Tooth-IAC evidence 8 findings Tooth-centered 3D crop Gated Fusion Findings Probability Whole Volume Appearance Encoder Global Token Findings Probability Left Posterior Crop Right posterior crop Findings Probability DISCRETE DENTAL EVIDENCE (E) TypeKeyPred.Conf. tooth 36missing0.92 tooth 46implant0.95 Global mandibular atrophy 0.88 48-IACclose0.86 ............ 11 12 13 14 15 16 17 18 21 22 23 24 25 26 27 28 31 32 33 34 35 36 37 38 41 42 43 44 45 46 47 48 Upper Jaw Lower Jaw FDI Numbering tooth 36 | missing tooth 46 | implant Global | mandibular atrophy 48-IAC | close ...... TEETH · Tooth #36 is missing. · Implant is present at #46. GLOBAL FINDINGS · There is mandibular atrophy. TOOTH-IAC RELATIONS · Tooth 48 is near the IAC. ...... No Images, No visual token Inputs Content traceable to evidence tuples Figure 1: Overview of EviDent-CBCT. During training, routine clinical reports are converted into source-linked reference evidence that preserves verbatim supporting spans; explicit labels supervise all branches; eligible tooth-level non-mentions provide reduced-weight negative supervision. At inference, Stage 1 extracts anatomical masks and structure tokens from the CBCT volume. Stage 2 combines tooth-centered, whole-volume, and posterior-mandibular views to predict tooth, global, and tooth-IAC findings, which are consolidated into a discrete evidence record. A dental-logic consistency projection reconciles incompatible evidence before Stage 3 deterministically renders a canonical draft, which serves as the sole case-specific input to an image-blind local Qwen rewriter. The rewriter is trained through GPT-based content-preserving distillation on drafts constructed from consistency-projected predicted evidence. Solid arrows indicate inference; orange arrows signify training-only supervision; and light-blue arrows denote reduced-weight supervision from eligible tooth-level non-mentions. Dental logical-F1, versus 0.371 ± 0.018 for the strongest controlled direct interface. It ranked second in the automated evaluation and third in blinded clinical comparison in the ODIN 2026 challenge. Our main contributions are: • We introduce an evidence-bottlenecked framework in which a discrete evidence record is the only case-specific input to an image-blind report generator, supporting trans- parent dental-logic correction and an auditable record of selected case-specific content. • We design anatomy-aligned evidence learning under non- exhaustive report supervision, using class-dependent reli- ability weights for tooth-level non-mentions and a metal- sensitive channel for dental intensity cues. • We provide controlled direct-generation comparisons across five visual interfaces, two report targets, and general-purpose and medically aligned language de- coders, together with three-run tooth-branch ablations and a reference-derived analysis of evidence coverage, prediction, and realization. 2 Related Work Medical report generation. Medical report generation has evolved from recurrent and joint disease-report models (Jing, Xie, and Xing 2018; Wang et al. 2018) to memory-based, retrieval-based, and region-guided architectures (Chen et al. 2020, 2021; Endo et al. 2021; Tanida et al. 2023). Recent sys- tems increasingly connect visual encoders to large language models through frozen-LLM adaptation, diagnosis prompts, knowledge augmentation, and flexible token fusion (Wang et al. 2023; Liu et al. 2024a; Jin et al. 2024; Li et al. 2024; Wang et al. 2025; Goswami, Subedi, and Chakraborty 2025; Shi et al. 2026). Most prior work targets chest radiographs, while CT2Rep extends direct generation to volumetric CT (Hamamci, Er, and Menze 2024). Despite these advances, report decoders generally remain conditioned on continuous visual features, retrieved text, or soft diagnostic guidance. Factuality and evidence interfaces. Factual correctness has consequently become a central concern. Prior work op- timizes clinical coherence and entity, relation, or semantic- level agreement (Liu et al. 2019; Miura et al. 2021; Delbrouck et al. 2022); RadGraph supports structured clinical extrac- tion (Jain et al. 2021), and RadFact evaluates reports through bidirectional proposition entailment (Bannur et al. 2024). Grounding, preference optimization, and reasoning further improve visual attribution and factuality (Bannur et al. 2024; Xiao et al. 2025; Jing et al. 2026), but do not require genera- tion to proceed through an explicit record of predicted find- ings. Structured approaches such as KERP, region-guided generation, and concept bottlenecks instead separate content prediction from verbalization (Li et al. 2019; Koh et al. 2020; Tanida et al. 2023). EviDent-CBCT makes this separation a hard inference-time constraint: an auditable evidence record is the sole case-specific input to an image-blind realizer. Dental CBCT Analysis. Dental CBCT learning has pri- marily focused on anatomical segmentation. nnU-Net pro- vides a strong general framework (Isensee et al. 2021); connectivity-aware methods address the thin inferior alveolar canal (Liu et al. 2024b; Bolelli et al. 2025); and recent work expands multi-structure segmentation through large dental benchmarks and state-space models (Bolelli et al. 2026; Tan et al. 2025). Moving toward clinical interpretation, Lumetti et al. formulate dental CBCT reporting as ontology-grounded prediction of finding presence, anatomical allocation, and clinical qualifiers from clinician-validated structured targets (Lumetti et al. 2026). Their system treats structured predic- tion as the final output and does not address narrative real- ization; its task-specific schema and structured metrics are therefore not directly comparable to free-text report genera- tion. EviDent-CBCT instead learns tooth-level, global, and tooth-IAC evidence from selective, non-exhaustive routine reports and makes the predicted discrete record the required case-specific input to an image-blind narrative realizer. 3 Problem Formulation Let x denote a 3D CBCT volume and y its observed clini- cal report. Direct generation estimates p(y | x). We instead derive a finite evidence record E y = (E y tooth ,E y global ,E y relation )(1) from the report.E y tooth contains findings indexed by FDI tooth position, E y global contains volume-level findings, and E y relation contains tooth-IAC and canal-course findings. The super- script emphasizes that this is evidence observed through y, not exhaustive image truth. Reporting is factorized as b E = q θ (x), e E = Π dent ( b E), y 0 = R( e E),by = G φ (y 0 ), (2) where q θ predicts evidence, Π dent is a deterministic dental- logic consistency projection, R is a deterministic renderer, and G φ is an image-blind rewriter. Every content item in y 0 is generated from a tuple in e E. The rewriter is instructed to preserve finding identity, polar- ity, and FDI number while altering ordering, aggregation, syntax, and lexical choice. Because learned realization is not guaranteed to be lossless, retainingy 0 makes rewrite-induced deviations directly auditable and allows evidence prediction to be evaluated separately from language realization. 4 Method Overview As illustrated in Figure 1, the method isolates these measur- able interfaces: anatomical localization, report-derived ev- idence recovery, deterministic consistency correction, and textual realization of the selected evidence. Stage 1 supplies anatomical masks and reusable structure tokens. Stage 2 maps anatomy-aligned inputs to a partially observed mul- tilabel record. Stage 3 converts the record to prose while keeping the content interface discrete. Reference Evidence Construction Before model training, we construct reference evidence from each observed report with a two-pass, dental-knowledge- guided workflow. The first pass GPT-5.5 agent identifies coverage statements, tooth-indexed findings, tooth-IAC rela- tions, global clinical facts, and normal or explicitly negative findings, while retaining the supporting report span for ev- ery item. A review-and-correction agent implemented with Claude Opus 4.8 then verifies each structured item against its source text, confirming 98.3% on the first pass and directly correcting the remaining 1.7% before Stage 2 label con- struction. Schema definitions and verification procedures are provided in Sect. B of supplementary. The resulting JSON is deliberately richer than the com- pact Stage 2 interface: it combines normalized entries with free-form facts and source text, whereas Stage 2 projects recurrent, learnable content into 11 tooth, 12 global, and 8 relation categories. The complete prediction schema and label frequencies are listed in supplementary Sect. B. The workflow structures the observed report; it does not claim to adjudicate the report or create exhaustive clinical truth. Stage 1: Anatomy Segmentation and Feature Exposure Stage 1 is an nnU-Net-v2-style 3D U-Mamba2 segmenter (Tan et al. 2025) trained on ToothFairy3 to predict 46 segmentation classes, including FDI-position tooth masks, mandible, maxilla, sinuses, canals, and dental-material masks. It supplies anatomical masks for crop construction and semantic channels, together with pooled encoder tokens for individual structures. The segmenter achieves a macro Dice of 0.899 over present structures and a case-averaged Dice of 0.880; training and feature-extraction details are pro- vided in Sect. C of supplementary. Stage 2: Reliability-Aware Evidence Prediction Branch-specific views. The schema is decomposed into three multilabel branches with separate anatomical inputs and classifiers. The tooth branch predicts 11 findings at each of 32 FDI positions from tooth-centered crops. The global branch predicts 12 volume-level findings from a down- sampled full volume. The relation branch predicts eight canal-related findings from left and right posterior-mandible views. This decomposition aligns each receptive field and output space with the corresponding clinical unit. Anatomy-aligned fusion. Each branch contains a from- scratch 3D encoder that produces an appearance feature r. The global branch uses r directly. The relation and tooth branches project the corresponding Stage 1 token into a con- text feature c and use an anatomical embedding s to identify the tooth or region. Their gated hybrid representation is g = σ(W g [r;c;s] +b g ), z = (1−g)⊙r +g⊙c. (3) whereW g andb g are the learnable projection matrix and bias of the element-wise fusion gate. The hybrid representation can exploit segmentation context while retaining raw inten- sity patterns that Stage 1 may discard. The main system uses a scratch global branch and a hybrid relation branch. The corresponding global-relation interface sweep is reported in supplementary Sect. F. Metal-sensitive intensity channel. Standard preprocess- ing compresses high-intensity CBCT values, obscuring dif- ferences among metallic restorations and prostheses. We therefore add a raw-intensity channel to each tooth crop, with its window estimated from high-density dental-material regions within the tooth-centered crop. It suppresses lower- density tissue while preserving normalized variation among high-density materials and artifacts, complementing the con- ventional anatomical channel. Reliability-aware tooth-level non-mention supervision. Report parsing assigns each tooth-class pair an observed positive, an observed negative, or an unmentioned state. We apply non-mention supervision only to eligible tooth- level labels; unreported global and tooth-IAC labels remain unknown and are excluded from the corresponding losses. Motivated by the hypothesis that eligible tooth-level non- mentions are predominantly negative in aggregate, while re- maining uncertain at the individual-case level, we compare three policies: explicit excludes it from the loss, hard assigns it a unit-weight negative label, andpu-weak assigns it a negative label with reduced, class-dependent weight. This reliability-weighting strategy is inspired by positive- unlabeled (PU) learning (Kiryo et al. 2017). The weighted tooth-branch objective is L = X i,c β c w(s ic ,c) BCE(ey ic ,p ic ),(4) where β c addresses class prevalence and w represents su- pervision reliability. Explicit positive and negative state- ments receive unit weight, whereas eligible tooth-level non- mentions receive 0 < w(unmentioned,c) < 1 under pu-weak. The schema-specific supervision rules and class- dependent weights are detailed in supplementary Sect. B. Tooth-level sampling and missingness conditioning. Within the tooth branch, per-class epoch sampling retains all supervised positives while limiting exposure to abun- dant negatives. A detached missing-tooth probability condi- tions the remaining attribute head, allowing it to distinguish present and likely-absent tooth contexts without propagating attribute loss into missingness prediction. The sampling and conditioning details are given in Sect. C of supplementary. Stage 3: Evidence-Bottlenecked Report Realization Dental-logic consistency projection. Before rendering, deterministic constraints derived from dental anatomical se- mantics map b E to e E = Π dent ( b E). Rather than treating se- lected labels independently, the projection removes combina- tions that violate predefined dental compatibility constraints, using predicted confidence only to resolve mutually exclu- sive alternatives. It operates solely on predicted tuples, in- troduces no positive finding, and leaves the corrected record inspectable at the evidence interface. The complete constraint definitions are provided in supplementary Sect. C. Deterministic evidence rendering. A non-parametric ren- derer converts the consistency-projected evidence set e E into a canonical draft. Findings are grouped by compatible anatomical region and finding type, with compact templates used for repetitive patterns such as a fully edentulous arch. The resulting draft explicitly enumerates the content selected by Stage 2 while leaving sentence structure, ordering, aggre- gation, and clinical phrasing to the learned realizer. Content-preserving teacher rewrites. For each training case, we construct the draft from consistency-projected pre- dicted evidence, thereby exposing the realizer to inputs re- sembling those encountered at deployment. GPT-4o (OpenAI 2024) then rewrites the draft into natural clinical language. The corresponding clinical report is provided only as a lin- guistic reference, and GPT-4o is instructed to preserve the factual content of the draft without adding, removing, or cor- recting findings. The resulting rewrites serve as realization targets rather than corrected evidence labels. Image-blind behavioral distillation. We distill the teacher’s evidence-to-report realization behavior into a lo- cal Qwen2.5-7B rewriter adapted with low-rank adaptation (LoRA) (Hu et al. 2022). The student receives only the de- terministic draft during both training and inference; it has no access to the CBCT volume, the reference report, or a parallel visual representation through which it could bypass the evidence bottleneck. It is therefore not explicitly trained to identify or correct Stage 2 prediction errors. We retain the deterministic draft for direct comparison, and inference re- quires no external model call. Decoding details are reported in supplementary Sect. C. We empirically evaluate the trained student rewriter’s content preservation and rewrite fidelity in Auditability and Rewrite Fidelity, with additional analyses provided in supplementary Sect. D. 5 Experiments Dataset The paired subset of ToothFairy4 contains 624 CBCT vol- umes and 1,001 English reports (Lumetti et al. 2026). Re- ports associated with the same volume are consolidated at EviDent-CBCT CBCT CBCT CBCT CBCT Seg-Pretrained Stage-1 Encoder (Frozen) Seg-Pretrained Stage-1 Encoder (Trainable) Evidence-Pretrained Stage-2 Encoder (Trainable) Evidence-Pretrained Stage-2 Encoder (Trainable) Stage 1 - Seg Stage 1 - Seg Stage 1 Anatomy-Seg Stage 2 Evidence Prediction Consistency-Projected Evidence Bottleneck Stage 3 Report Generation Feature Adaptor Feature Adaptor Feature Adaptor Feature Adaptor Report LLM Decoder LLM Decoder LLM Decoder LLM Decoder Report Report Report Auxiliary Evidence Head (Training Only) Report Deterministic Renderer Image-blind LLM Rewriter Direct-1: Frozen Stage1 Direct-2: Fine-tuned Stage1 Direct CBCT-to-Report Generation Direct-3: Continuous Stage-2 Features Direct-4: Direct 3 + Auxiliary Evidence Evidence Teeth: Findings: Caries, Root Fracture 36, 37 TypeKeyPred.Conf. tooth 36 missing 0.92 tooth 46 implant 0.95 Global mandibular atrophy 0.88 48-IAC close 0.86 ............ LLM Decoder Report Evidence-Pretrained Stage-2 Encoder (Frozen) Stage 1 - Seg Feature Adaptor Direct-5: Frozen Stage-2 Features CBCT Figure 2: Comparison of the inference-time interfaces used by the controlled direct generators and EviDent-CBCT. Direct 1 and Direct 2 condition a language decoder on frozen or fine-tuned Stage 1 representations. Direct 3 fine-tunes Stage 2- checkpoint-initialized branches and supplies their continuous features to the decoder; Direct 4 adds training-only auxiliary evidence supervision, while Direct 5 freezes the Stage 2 branches. These five interfaces are instantiated with Qwen3-8B and Med42-8B decoders. EviDent-CBCT instead supplies a consistency-projected evidence record as the only case-specific input to an image-blind realizer. Its deterministic draft exposes the predicted findings and makes rewrite-induced deviations auditable. the label level; when they disagree on a binary finding, we exclude that label from supervision rather than arbitrate be- tween readers. As each scan contributes many tooth-level instances, this conservative treatment still provides ample, unambiguous supervision for Stage 2, while modeling inter- report disagreement remains future work. We split the 624 paired volumes into 561 training and 63 validation cases. The same volume-level partition was applied to all three stages and all direct baselines. Any ToothFairy3 volume cor- responding to a validation case was excluded from Stage 1 training. ToothFairy4 additionally includes a hidden test set of 50 CBCT scans from an independent center, evaluated by the challenge organizers; official results are reported below. Evaluation Metrics Merged Stage 2 logical set-F1. For each case, positive tu- ples parsed from the reference report are pooled across the tooth, global, and relation branches within the evaluation mask. Predicted and reference-derived tuples match only when their structured keys are identical. We compute per- case precision and recall, average each quantity over cases, and take the harmonic mean of those two averages. This RadFact-style aggregation is an LLM-free measure of struc- tured agreement with the reports. Merged Stage 2 logical set-F1 is computed on the selected tuples before consistency projection, isolating the learned evidence predictor from de- terministic correction. Its formal definition and aggregation conventions are given in supplementary Sect. G. Internal report metrics. RadFact decomposes each narra- tive report into clinical propositions and evaluates entailment bidirectionally (Bannur et al. 2024). Candidate propositions supported by the reference determine logical precision, while reference propositions supported by the candidate determine logical recall; logical F1 is their harmonic mean. We build on RadFact-Lite (Lumetti 2026) with a dentist-guided dental- atomic parser, FDI normalization, and dental entailment in- structions. Exact and rule-guaranteed matches are resolved deterministically, while comparisons are judged by GPT-5.6- Luna. The original bidirectional aggregation is unchanged. We report these metrics together with BLEU-4 (Papineni et al. 2002) and METEOR (Banerjee and Lavie 2005). Main- paper metrics use the same fixed reference report per volume; supplementary Sect. E provides a mean-over-references ro- bustness analysis. Full details and an author audit of RadFact- Lite-Dental are provided in supplementary Sect. G. The of- ficial ODIN evaluation uses a different RadFact protocol. We therefore report its scores separately and do not directly compare their absolute values with RadFact-Lite-Dental. Paired uncertainty analysis. For the primary comparison, we average case-level logical metrics across runs, jointly bootstrap matched cases, recompute corpus-level logical F1, and report the percentile 95% confidence interval of paired differences. Details are provided in supplementary Sect. E. EviDent-CBCT vs. Controlled Direct Generation Figure 2 summarizes EviDent-CBCT and five controlled direct-generation interfaces. Direct 1 projects frozen Stage 1 FamilyVisual/report interfaceReport targetLogical PLogical RLogical F1BLEU-4METEOR Direct Qwen3-8B Direct 1: frozen Stage 1 tokensRaw.216± .016 .245± .036 .229± .018 .121± .005 .352± .004 Direct 1: frozen Stage 1 tokensCanonical.295± .033 .294± .043 .294± .034 .095± .004 .349± .006 Direct 2: fine-tuned Stage 1 encoderRaw.257± .027 .251± .009 .253± .011 .135± .007 .338± .012 Direct 2: fine-tuned Stage 1 encoderCanonical.331± .056 .315± .023 .322± .036 .093± .007 .355± .005 Direct 3: continuous Stage 2 featuresRaw.238± .034 .236± .030 .236± .025 .124± .015 .325± .024 Direct 3: continuous Stage 2 featuresCanonical.351± .004 .314± .009 .331± .005 .099± .003 .356± .005 Direct 4: Direct 3 + auxiliary evidence Raw.240± .038 .249± .030 .244± .034 .119± .015 .322± .007 Direct 4: Direct 3 + auxiliary evidence Canonical.342± .008 .322± .027 .331± .014 .098± .003 .357± .015 Direct 5: frozen Stage 2 encoderRaw.239± .035 .233± .017 .234± .010 .121± .014 .326± .015 Direct 5: frozen Stage 2 encoderCanonical.346± .018 .311± .020 .327± .003 .097± .002 .358± .011 Direct Med42-8B Direct 1: frozen Stage 1 tokensRaw.227± .016 .264± .015 .243± .005 .121± .015 .360± .006 Direct 1: frozen Stage 1 tokensCanonical.356± .020 .334± .014 .344± .012 .098± .002 .353± .007 Direct 2: fine-tuned Stage 1 encoderRaw.266± .031 .293± .016 .278± .024 .129± .009 .352± .009 Direct 2: fine-tuned Stage 1 encoderCanonical.388± .019 .354± .017 .371± .018 .100± .002 .360± .003 Direct 3: continuous Stage 2 featuresRaw.291± .003 .272± .036 .280± .018 .134± .012 .341± .011 Direct 3: continuous Stage 2 featuresCanonical.366± .005 .359± .011 .362± .006 .099± .001 .357± .003 Direct 4: Direct 3 + auxiliary evidence Raw.273± .020 .272± .019 .271± .002 .127± .017 .324± .017 Direct 4: Direct 3 + auxiliary evidence Canonical.374± .008 .356± .010 .365± .007 .101± .002 .361± .004 Direct 5: frozen Stage 2 encoderRaw.252± .004 .274± .020 .262± .010 .120± .013 .313± .011 Direct 5: frozen Stage 2 encoderCanonical.380± .024 .342± .006 .360± .013 .099± .001 .354± .003 EviDent-CBCTConsistency-projected evidence record Deterministic template .372± .002 .428± .008 .398± .004 .158± .003 .425± .004 Consistency-projected evidence record Local distilled rewrite .384± .002 .423± .004 .402± .003 .083± .001 .386± .001 Table 1: Validation set results, reported as mean±std over three runs. Logical metrics use RadFact-Lite-Dental. structure tokens, while Direct 2 recomputes these tokens from a fixed-size volume and fine-tunes the encoder lev- els that produce them. Direct 3 initializes the global, rela- tion, and tooth branches from the corresponding Stage 2 checkpoints and supplies their trainable continuous fea- tures to the language decoder; Direct 4 additionally ap- plies Stage 2 matched auxiliary supervision; and Direct 5 freezes the Stage 2 branches while retaining a trainable fea- ture adapter and decoder, testing whether direct SFT de- grades the evidence-pretrained representation. Each interface is instantiated with either Qwen3-8B (Yang et al. 2025) or Med42-8B (Christophe et al. 2024) and trained with both raw and canonical report targets. The raw target is the observed clinical report, whereas the canonical target is obtained by rendering reference-derived evidence and passing the draft through the frozen rewriter, thereby reducing linguistic and content-selection variability. Further implementation details and qualitative comparisons are provided in supplementary Sect. E and Sect. I, respectively. Canonical targets improve every decoder-interface pair. Med42 strengthens the direct controls, with canonical Direct 2 reaching .371± .018 F1, compared with at most .331 for Qwen3. Among Stage 2-initialized Med42 controls, Direct 3 and Direct 4 reach .362±.006 and .365±.007, respectively. Freezing the Stage 2 branches in Direct 5 yields a decline with both Qwen3 and Med42, providing no improvement over the corresponding trainable Direct 3 pathways. EviDent-CBCT reaches .398± .004 logical-F1 with de- terministic templates and .402± .003 after local rewriting. The template also obtains the highest BLEU-4 and ME- TEOR, showing that the comparison does not depend on the learned realizer. Against the strongest Med42 direct con- figuration, the rewritten output has an unrounded paired F1 gain of .0317, with a 95% bootstrap confidence interval of [.0047, .0586]. Moreover, Direct 3 and Direct 4 use the same evidence-pretrained Stage 2 initialization while retain- ing a continuous, bypassable interface. These controls sup- port making the discrete evidence record the required report interface rather than only a pretraining signal or auxiliary tar- get. Per-run results and specifications are reported in supple- mentary Sect. E. This advantage persists under mean-over- references evaluation (.374 vs. .333; supplementary Sect. E). Stage-wise Analysis of Evidence-to-Report Pathway Table 2 traces the successive representations from the observed report to the generated output. Reference self- comparison reaches 1.000 logical-F1. The extracted source spans cover 96.5% of alphanumeric report content on aver- age, and their verbatim reassembly reaches .957 logical-F1. Together with the second-agent verification pass, this indi- cates that the extraction workflow retains most evaluator- recognized report content. The residual consists mainly of sparse orientation notes, classification terms, and broad nor- mal statements; detailed coverage analysis is provided in supplementary Sect. B. The largest loss occurs when the richer evidence JSON is projected into the compact prediction schema and deter- ministically rendered, reducing logical-F1 from .957 to .589. The schema preserves recurrent finding identities but omits details without dedicated labels, including the distribution and degree of atrophy, treatment quality, bridge topology, and lesion morphology or size. Qwen realization raises the reference-derived score to .597 by restoring narrative organi- zation, aggregation, and clinically familiar phrasing. Because Candidate reportLogical F1 Reference report vs. itself1.000 Verbatim source clues from evidence extraction.957 Reference-derived evidence→ template.589 Reference-derived evidence→ template→ Qwen.597 Predicted evidence→ template.389± .006 Predicted evidence→ template→ Qwen.398± .001 Consistency-projected evidence→ template.398± .004 Consistency-projected evidence→ template→ Qwen.402± .003 Table 2: Evidence-to-report pathway decomposition and reference-derived operational ceiling. Reference-derived rows are evaluated once; predicted-evidence rows are re- ported as mean±std over three runs. AxisVariantLogical PLogical RLogical F1 Basegated, metal, pu-weak .434± .028 .538± .022 .480± .007 Supervision hard.408± .017 .548± .030 .467± .017 explicit.405± .057 .328± .035 .359± .018 Fusionconcat.435± .036 .525± .033 .474± .015 visual-only.422± .018 .534± .033 .471± .004 scratch-only.379± .021 .512± .039 .435± .001 Metalno metal channel.415± .019 .527± .039 .463± .006 Table 3: Tooth-branch Stage 2 logical set metrics, reported as mean±std over three model runs. All non-ablated branches and inputs are held fixed. the rewriter receives only the deterministic draft, .597 is an operational ceiling for the current compact schema-to-report pathway, not for the richer evidence JSON. We next quantify the effect of the Stage 3 dental-logic consistency projection. Its constraints were fixed before the final report-level evaluation and modified 178 tuple in- stances across 52 unique validation cases over three runs: 162 missing-tooth/IAC conflicts and 16 contradictory canal- course predictions. The projection raises template F1 from .389±.006 to .398±.004 and rewritten F1 from .398±.001 to.402±.003, showing that correction occurs at the evidence interface rather than in the language realizer. The remaining gap to the reference-derived pathway primarily reflects up- stream evidence prediction. Auditability and Rewrite Fidelity The deterministic draft makes learned realization auditable at the proposition level. An author reviewed 30 evalua- tion cases for additions, omissions, polarity changes, FDI or anatomical-side reassignment, finding substitutions, and redundancy. All template propositions were preserved in 28 cases (93.3%), with no polarity reversal, FDI reassign- ment, or anatomical-side reassignment. The two substantive changes both aligned the rewrite more closely with the ref- erence: one removed a crown prediction where the reference described a restoration, and the other changed a restoration to a crown. Rewriting is therefore not guaranteed to be loss- less, even though individual changes can regularize imperfect predicted evidence. Its modest aggregate improvement is re- ported in Table 2, and retaining the pre-rewrite draft keeps case-level changes inspectable. The full audit protocol and case-level deviations are reported in supplementary Sect. D. Stage 2 Evidence Learning and Tooth Ablations The final g_scratch_r_hybrid pipeline uses a scratch global branch, a hybrid relation branch, and the tooth-base configuration. Across three runs, it reaches .666 ± .006 merged Stage 2 logical set-F1 and .402± .003 final report logical-F1. Tooth evidence is the most fine-grained compo- nent of the merged record and the focus of Table 3. Sup- plementary Sect. H further reports class-wise recognition results for the tooth branch under pu-weak, hard, and explicit supervision, together with class-wise results for the fixed global and relation branches. Within the tooth branch, treating every eligible non- mention as a full-weight negative (hard) raises logical set-F1 from .359 ± .018 with explicit supervision to .467±.017, primarily by increasing recall from .328 to .548. This suggests that tooth-level report silence provides useful negative signal in aggregate. Assigning class-dependent re- duced weights to these non-mentions (pu-weak) gives the best balance, with .434±.028 precision, .538±.022 recall, and .480± .007 F1. Within the current split and weighting scheme, these results support reduced-weight supervision for eligible tooth-level non-mentions, but do not establish that the selected weights are clinically calibrated or that the same policy applies to global and relation labels. The fusion variants are close: gated hybrid, concat, and visual-only reach tooth-branch F1 values of .480, .474, and .471, whereas scratch-only falls to.435. Removing the metal- sensitive channel lowers tooth-branch F1 from .480 to .463. These ablations are measured directly at the Stage 2 tooth interface, with the global and relation branches held fixed. Official ODIN 2026 Hidden-Test Evaluation On the ODIN 2026 hidden test set 1 , EviDent-CBCT achieved an official aggregate score of 0.3033 (RadFact F1 0.3245, BLEU-4 0.1206, METEOR 0.3171), ranking second in auto- mated evaluation and third in the clinical Arena. The official RadFact protocol differs from our RadFact-Lite-Dental eval- uation, so the absolute scores are not directly comparable. 6 Conclusion and Outlook We presented EviDent-CBCT, an evidence-bottlenecked framework for learning dento-maxillofacial CBCT reports from non-exhaustive routine-report supervision. Its discrete record separates evidence prediction from language real- ization, enables deterministic dental-consistency correction, and makes selected report content auditable. For tooth-level evidence, eligible non-mentions receive reduced-weight neg- ative supervision, while unreported global and tooth-IAC labels remain unknown. Across three validation runs, the system achieves 0.402± 0.003 RadFact-Lite-Dental logical- F1, compared with 0.371± 0.018 for the strongest controlled direct generator. It further ranked second in the ODIN 2026 1 https://odin2026.grand-challenge.org/challenge-winners/ Table 4: Official ODIN 2026 ToothFairy4 hidden-test results for the top three teams. The official RadFact protocol differs from RadFact-Lite-Dental used in our internal validation. TeamRadFact F1 BLEU-4 METEOR Aggregate Score Arena Score shivam.singh0.39240.14760.32690.36141307 EviDent-CBCT(ours)0.32450.12060.31710.3033985 ViktorB0.23210.08650.28280.22261172 automated evaluation and third in the blinded clinical com- parison. The controlled comparisons, supervision ablations, and coverage analysis support the discrete evidence record as a structured content interface for report-grounded CBCT generation. These findings remain limited by non-exhaustive report supervision, the lack of independently adjudicated image- level truth, and a compact schema that underrepresents rare qualifiers and findings. The hidden-test evaluation provides independent-center testing and blinded clinical assessment, although the test references remain unavailable for detailed error analysis. Future work will model cross-report disagree- ment, preserve rare qualifiers, and evaluate additional exter- nal cohorts with blinded dental readers. References Banerjee, S.; and Lavie, A. 2005. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judg- ments. In Proceedings of the ACL Workshop on Intrinsic and Ex- trinsic Evaluation Measures for Machine Translation and/or Sum- marization, 65–72. Ann Arbor, Michigan: Association for Compu- tational Linguistics. Bannur, S.; Bouzid, K.; Castro, D. C.; Schwaighofer, A.; Thieme, A.; Bond-Taylor, S.; Ilse, M.; Pérez-García, F.; Salvatelli, V.; Sharma, H.; et al. 2024. MAIRA-2: Grounded radiology report generation. arXiv:2406.04449. Bannur, S.; Hyland, S.; Liu, Q.; Perez-Garcia, F.; Ilse, M.; Castro, D. C.; Boecking, B.; Sharma, H.; Bouzid, K.; Thieme, A.; et al. 2023. Learning to exploit temporal structure for biomedical vision- language processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15016–15027. Van- couver, British Columbia, Canada: IEEE. Bolelli, F.; Lumetti, L.; van Nistelrooij, N.; Vinayahalingam, S.; Di Bartolomeo, M.; Marchesini, K.; Pellacani, A.; Candeloro, E.; Rosati, G.; Xi, T.; et al. 2026. Multi-structure segmentation in CBCT volumes: The ToothFairy2 challenge. Medical Image Analysis, 112: 104095. Bolelli, F.; Lumetti, L.; Vinayahalingam, S.; Di Bartolomeo, M.; Pellacani, A.; Marchesini, K.; van Nistelrooij, N.; van Lierop, P.; Xi, T.; Liu, Y.; et al. 2025. Segmenting the inferior alveolar canal in CBCTs volumes: The ToothFairy challenge. IEEE Transactions on Medical Imaging, 44(4): 1890–1906. Chen, Z.; Shen, Y.; Song, Y.; and Wan, X. 2021. Cross-modal mem- ory networks for radiology report generation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Pro- cessing, 5904–5914. Online: Association for Computational Lin- guistics. Chen, Z.; Song, Y.; Chang, T.-H.; and Wan, X. 2020. Generating radiology reports via memory-driven transformer. In Proceedings of the Conference on Empirical Methods in Natural Language Pro- cessing, 1439–1449. Online: Association for Computational Lin- guistics. Christophe, C.; Kanithi, P. K.; Raha, T.; Khan, S.; and Pimentel, M. A. F. 2024. Med42-v2: A suite of clinical LLMs. arXiv:2408.06142. Delbrouck, J.-B.; Chambon, P.; Bluethgen, C.; Tsai, E.; Almusa, O.; and Langlotz, C. 2022. Improving the factual correctness of radiology report generation with semantic rewards. In Proceedings of the Findings of the Association for Computational Linguistics: Conference on Empirical Methods in Natural Language Processing, 4348–4360. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics. Delbrouck, J.-B.; Xu, J.; Moll, J.; Thomas, A.; Chen, Z.; Ostmeier, S.; Azhar, A.; Li, K. Z.; Johnston, A.; Bluethgen, C.; et al. 2025. Au- tomated structured radiology report generation. In Proceedings of the Annual Meeting of the Association for Computational Linguis- tics, 26813–26829. Vienna, Austria: Association for Computational Linguistics. Endo, M.; Krishnan, R.; Krishna, V.; Ng, A. Y.; and Rajpurkar, P. 2021. Retrieval-based chest X-ray report generation using a pre-trained contrastive language-image model. In Proceedings of Machine Learning for Health, 209–219. Virtual: PMLR. Goswami, D.; Subedi, R.; and Chakraborty, S. 2025. MediVLM: A Vision Language Model for Radiology Report Generation from Medical Images. In Proceedings of the Findings of the Association for Computational Linguistics: Conference on Empirical Methods in Natural Language Processing, 10287–10304. Suzhou, China: Association for Computational Linguistics. Hamamci, I. E.; Er, S.; and Menze, B. 2024. CT2Rep: Automated radiology report generation for 3D medical imaging. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, 476–486. Marrakesh, Morocco: Springer Nature Switzerland. Hartung, M. P.; Bickle, I. C.; Gaillard, F.; and Kanne, J. P. 2020. How to create a great radiology report. RadioGraphics, 40(6): 1658–1670. Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations. Virtual. Isensee, F.; Jaeger, P. F.; Kohl, S. A. A.; Petersen, J.; and Maier- Hein, K. H. 2021. nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2): 203–211. Jain, S.; Agrawal, A.; Saporta, A.; Truong, S. Q. H.; Duong, D. N.; Bui, T.; Chambon, P.; Zhang, Y.; Lungren, M. P.; Ng, A. Y.; Lan- glotz, C. P.; and Rajpurkar, P. 2021. RadGraph: Extracting clinical entities and relations from radiology reports. In Proceedings of the Conference on Neural Information Processing Systems, Datasets and Benchmarks Track, volume 1. Virtual: Curran Associates, Inc. Jin, H.; Che, H.; Lin, Y.; and Chen, H. 2024. PromptMRG: Diagnosis-driven prompts for medical report generation. In Pro- ceedings of the AAAI Conference on Artificial Intelligence, 3, 2607– 2615. Vancouver, British Columbia, Canada: AAAI Press. Jing, B.; Xie, P.; and Xing, E. 2018. On the automatic generation of medical imaging reports. In Proceedings of the Annual Meet- ing of the Association for Computational Linguistics, 2577–2586. Melbourne, Australia: Association for Computational Linguistics. Jing, P.; Lee, K.; Zhang, Z.; Zhou, H.; Yuan, Z.; Gao, Z.; Zhu, L.; Papanastasiou, G.; Fang, Y.; and Yang, G. 2026. Reason like a radi- ologist: Chain-of-thought and reinforcement learning for verifiable report generation. Medical Image Analysis, 109: 103910. Kaasalainen, T.; Ekholm, M.; Siiskonen, T.; and Kortesniemi, M. 2021. Dental cone beam CT: An updated review. Physica Medica, 88: 193–217. Kiryo, R.; Niu, G.; du Plessis, M. C.; and Sugiyama, M. 2017. Positive-unlabeled learning with non-negative risk estimator. In Proceedings of the Conference on Neural Information Processing Systems, volume 30, 1675–1685. Long Beach, California, USA: Curran Associates, Inc. Koh, P. W.; Nguyen, T.; Tang, Y. S.; Mussmann, S.; Pierson, E.; Kim, B.; and Liang, P. 2020. Concept bottleneck models. In Proceedings of the International Conference on Machine Learning, 5338–5348. Virtual: PMLR. Li, C. Y.; Liang, X.; Hu, Z.; and Xing, E. P. 2019. Knowledge-driven encode, retrieve, paraphrase for medical image report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, 1, 6666–6673. Honolulu, Hawaii, USA: AAAI Press. Li, Y.; Liu, Y.; Wang, Z.; Liang, X.; Liu, L.; Wang, L.; and Zhou, L. 2025. S-RRG-Bench: Structured radiology report generation with fine-grained evaluation framework. Meta-Radiology, 3(4): 100171. Li, Y.; Wang, Z.; Liu, Y.; Wang, L.; Liu, L.; and Zhou, L. 2024. KARGEN: Knowledge-enhanced automated radiology report gen- eration using large language models. In Proceedings of the Inter- national Conference on Medical Image Computing and Computer- Assisted Intervention, 382–392. Marrakesh, Morocco: Springer Na- ture Switzerland. Liu, C.; Tian, Y.; Chen, W.; Song, Y.; and Zhang, Y. 2024a. Boot- strapping large language models for radiology report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, 17, 18635–18643. Vancouver, British Columbia, Canada: AAAI Press. Liu, G.; Hsu, T.-M. H.; McDermott, M.; Boag, W.; Weng, W.-H.; Szolovits, P.; and Ghassemi, M. 2019. Clinically accurate chest X-ray report generation. In Proceedings of the Machine Learning for Healthcare Conference, 249–269. Ann Arbor, Michigan, USA: PMLR. Liu, Y.; Xin, R.; Yang, T.; and Wang, L. 2024b. Inferior alveolar nerve segmentation in CBCT images using connectivity-based se- lective re-training. In Proceedings of the Challenges on Supervised and Semi-Supervised Multi-Structure Segmentation and Landmark Detection in Dental Data, Held in Conjunction with the Interna- tional Conference on Medical Image Computing and Computer- Assisted Intervention, 3–12. Marrakesh, Morocco: Springer Nature Switzerland. Lumetti, L. 2026. RadFact-Lite, version 0.1.0. Python package. Accessed June 20, 2026. Lumetti, L.; Di Bartolomeo, M.; Pellacani, A.; Anesi, A.; Grana, C.; and Bolelli, F. 2026. Ontology-grounded structured prediction for dental CBCT reporting. In Proceedings of the International Con- ference on Medical Image Computing and Computer-Assisted Inter- vention, 408–417. Strasbourg, France: Springer Nature Switzerland. Miura, Y.; Zhang, Y.; Tsai, E.; Langlotz, C.; and Jurafsky, D. 2021. Improving factual completeness and consistency of image-to-text radiology report generation. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 5288–5304. Online: Association for Computational Linguistics. OpenAI. 2024. GPT-4o system card. arXiv:2410.21276. Pandharipande, P. V.; Herts, B. R.; Gore, R. M.; Mayo-Smith, W. W.; Harvey, H. B.; Megibow, A. J.; and Berland, L. L. 2016. Rethinking normal: Benefits and risks of not reporting harmless incidental findings. Journal of the American College of Radiology, 13(7): 764–767. Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceed- ings of the Annual Meeting of the Association for Computational Linguistics, 311–318. Philadelphia, Pennsylvania, USA: Associa- tion for Computational Linguistics. Reddy, K. S.; Guttal, K. S.; Nandimath, K.; and Voora, T. 2024. In- tegration of cone beam computed tomography for interdisciplinary diagnosis and treatment planning. Journal of Orofacial and Health Sciences, 11(4): 186–191. Shi, P.; Zhang, M.; Song, K.; Liu, J.; Gu, Y.; and Zhang, X. 2026. U-VLM: Hierarchical Vision Language Modeling for Report Gen- eration. arXiv:2603.00479. Tan, Z. Q.; Zhu, X.; Addison, O.; and Li, Y. 2025. U-Mamba2: Scal- ing state space models for dental anatomy segmentation in CBCT. In Proceedings of the International Workshop on Oral and Dental Image Analysis and Associated Challenges, Held in Conjunction with the International Conference on Medical Image Computing and Computer-Assisted Intervention, 122–132. Daejeon, South Ko- rea: Springer Nature Switzerland. Tanida, T.; Müller, P.; Kaissis, G.; and Rueckert, D. 2023. Inter- active and explainable region-guided radiology report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7433–7442. Vancouver, British Columbia, Canada: IEEE. Vosshenrich, J.; Nesic, I.; Boll, D. T.; and Heye, T. 2023. Investigat- ing the impact of structured reporting on the linguistic standardiza- tion of radiology reports through natural language processing over a 10-year period. European Radiology, 33(11): 7496–7506. Wang, X.; Peng, Y.; Lu, L.; Lu, Z.; Bagheri, M.; and Summers, R. M. 2018. TieNet: Text-image embedding network for common thorax disease classification and reporting in chest X-rays. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9049–9058. Salt Lake City, Utah, USA: IEEE. Wang, Z.; Liu, L.; Wang, L.; and Zhou, L. 2023. R2GenGPT: Radiology report generation with frozen LLMs. Meta-Radiology, 1(3): 100033. Wang, Z.; Sun, Y.; Li, Z.; Yang, X.; Chen, F.; and Liao, H. 2025. LLM-RG4: Flexible and factual radiology report generation across diverse input contexts. In Proceedings of the AAAI Conference on Artificial Intelligence, 8, 8250–8258. Philadelphia, Pennsylvania, USA: AAAI Press. Xiao, T.; Shi, L.; Liu, P.; Wang, Z.; and Bai, C. 2025. Radiology report generation via multi-objective preference optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, 8, 8664–8672. Philadelphia, Pennsylvania, USA: AAAI Press. Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv:2505.09388. Supplementary Material of EviDent-CBCT A Scope of the Report-Derived Supervision EviDent-CBCT starts from the clinical reports available in Tooth- Fairy4. We do not interpret these reports as exhaustive image an- notations or construct an independently adjudicated medical gold standard. A non-mention may reflect true absence, report-specific omission, habitual non-reporting, oversight, or disagreement across readers. These mechanisms are not adjudicated and cannot be dis- tinguished from the observed reports alone. Explicit report inaccu- racies are also possible, but their prevalence is not measured in this study. The three policies in the main paper apply only to eligible tooth- level non-mentions. explicit excludes these entries from the loss, hard assigns them unit negative weight, and pu-weak as- signs class-dependent reduced negative weights. Unreported global and tooth-IAC categories remain unknown and are excluded from supervision because their positive labels are sparse. None of these policies adjudicates explicit report inaccuracies, and our experi- ments do not estimate their prevalence. Accordingly, the reported Stage 2 metric measures structured agreement with the reports rather than clinical accuracy against an external standard. B Reference Evidence Construction and Coverage Audit Construction and independent verification The full reference evidence JSON is produced before Stage 2 pre- processing. GPT-5.5 (medium effort) receives the original report and a dental-specific extraction prompt covering FDI notation, arch and quadrant expressions, missing and treated teeth, prostheses and implants, bone findings, scan coverage, and tooth-IAC relations. It returns structured sections for coverage, tooth labels, relation labels, additional clinical facts, and normal or explicitly negative statements. Every item includes the report text that supports it, so the normalized entry remains traceable to the observed wording. A second agent implemented through Claude Opus4.8 indepen- dently reviews each JSON against the report. The reviewer checks support, polarity, tooth identity, anatomical assignment, complete- ness, and the correspondence of each attached span; it either accepts the entry or requests calibration. The reported 98.3% first-pass con- firmation rate is the proportion of entries that this Opus4.8 reviewer accepted without requesting calibration. The remaining 1.7% are flagged and corrected before constructing Stage 2 labels. This is agent-review agreement with the report, not dentist validation or correctness against an independently read CBCT volume. Full JSON versus the prediction schema Two representations should not be conflated. The full evidence JSON preserves categorical fields, free-form facts, normal/negative statements, and verbatim source evidence. The Stage 2 interface is a task-oriented projection containing 11 tooth-level, 12 global, and eight relation categories. Rare details can therefore remain auditable in the full JSON without becoming dedicated output classes for image-based prediction. The ceiling row “verbatim source clues” evaluates the former; the template and Qwen ceiling rows evaluate the latter after projection and realization. Complete Stage 2 evidence schema Tables 5-7 enumerate the complete prediction schema. Counts are positive reference-derived labels pooled across the fixed train- ing and validation partitions. Tooth counts denote FDI-position- category instances; global and relation counts denote case-category instances. They are label frequencies in the observed reports, not es- timates of clinical prevalence. For sparse global and relation labels, an unreported category remains unknown. For tooth labels, non- mentions eligible for positive-unlabeled supervision are assigned the reliability weights in Table 5; explicit positive and negative statements receive weight 1, while unavailable positions and con- tradictory statements receive weight 0. Coverage protocol For each evaluation case, we recursively collect the attached text fields from the coverage, tooth-label, relation-label, free-form fact, and normal/negative sections. After whitespace normalization, du- plicate clues and shorter clues wholly contained in longer ones are removed, and the retained spans are restored to their order in the original report. This yields 555 unique clues. Every retained clue is an exact substring after whitespace normalization. We report two complementary coverage measures. Literal cov- erage is the fraction of alphanumeric report characters contained in the union of retained source spans; its case-averaged value is 96.5% and its median is 100%. Semantic coverage concatenates the retained clues and compares them with the report using RadFact- Lite-Dental. The resulting logical precision, recall, and F1 are 0.986, 0.929, and 0.957, respectively. Literal coverage measures retained wording, whereas logical recall measures whether the selected spans support the propositions recovered from the report. Residual content We inspect report text outside the union of all retained source spans after excluding section headers, punctuation, and connective fragments. Table 8 summarizes the substantive residuals. They are concentrated in low-frequency metadata, classifications, and broad normal statements rather than recurring core dental findings. The audit supports a precise interpretation: major report con- tent is comprehensively captured, and the supporting text preserves most fine detail verbatim, but the normalized structured fields inten- tionally emphasize macroscopic, recurrent concepts. The residual concepts are too sparse in this corpus to justify separate super- vised classes. The decrease from 0.957 for source clues to 0.597 for reference-derived evidence after template rendering and Qwen rewriting should therefore be attributed primarily to post-extraction projection and realization, rather than to wholesale failure of the initial evidence extraction. Coverage-to-ceiling interpretation The coverage audit and ceiling study examine successive repre- sentations of the same observed reports. Source spans retained in the full evidence JSON preserve fine-grained statements verbatim, reaching 96.5% mean literal coverage and 0.957 logical-F1 when reassembled. Together with the two-pass verification results, this indicates that the extraction workflow captures the major content of the observed reports. The largest decrease occurs when this rich representation is pro- jected into the compact 31-category prediction schema and deter- ministically rendered, reducing logical-F1 from 0.957 to 0.589. This projection preserves recurrent findings but intentionally collapses details that are not modeled as dedicated labels, including the pre- cise distribution and degree of atrophy, treatment quality, bridge topology, and lesion morphology or size. Rare free-form findings and normal statements outside the schema are likewise unavail- able to the renderer. The decrease therefore reflects the combined Schema fieldMeaningTotal n + Observed w − Report-level w − missingTooth absent at the indexed FDI position924.95.95 implantEndosseous implant at the indexed position256.80.50 crownProsthetic crown at the indexed position695.60.40 restorationCoronal restoration or filling, distinct from a crown 1,792.50.30 endodontic_treatmentRoot-canal or other endodontic treatment766.70.40 prosthesis_relatedBridge, pontic, or other prosthetic involvement1,305.50.30 impacted_or_ disodontiasis Impaction or disodontiasis365.80.50 cariesDental caries233.20.10 root_remnantRetained root or root remnant61.70.40 root_fractureRoot fracture11.20.10 lesion_relatedTooth-associated lesion finding388.30.15 Table 5: Complete tooth-level prediction schema, positive sample counts, and class-dependent pu-weak weights. Observed denotes a non-mention for a tooth observed elsewhere in the report; report-level denotes an otherwise eligible non-mention inferred only from report silence. Schema fieldMeaningTotal n + maxilla_partially_includedMaxilla only partly included in the scan field236 maxilla_not_includedMaxilla excluded from the scan field234 maxillary_sinuses_not_ included Maxillary sinuses excluded from the scan field173 mandibular_condyles_not_ included Mandibular condyles excluded from the scan field588 periodontitis_presentReported periodontal disease207 atrophy_maxillaMaxillary alveolar atrophy53 atrophy_mandibleMandibular alveolar atrophy312 atrophy_severeSevere or marked atrophy in either arch113 osteolytic_lesion_presentOsteolytic or radiolucent osseous lesion200 osteosclerotic_lesion_present Osteosclerotic or osteocondensing lesion38 sinus_pathology_presentPathological finding in a maxillary sinus58 poor_quality_or_artifactPoor image quality or acquisition/reconstruction artifact22 Table 6: Complete global prediction schema and positive sample counts. granularity of the prediction schema and its deterministic realiza- tion, rather than a failure to identify those statements during initial extraction. Qwen realization raises logical-F1 from 0.589 to 0.597 for reference-derived evidence by restoring narrative organization, ag- gregation, and clinically familiar phrasing. Because the rewriter receives only the deterministic draft, it cannot systematically re- cover information removed during schema projection. The 0.597 reference-derived result is therefore an operational ceiling for the compact schema-to-report pathway, not for the richer evidence JSON. Replacing reference-derived evidence with predicted ev- idence lowers the rewritten score to 0.398 ± 0.001, isolating the remaining effect of Stage 2 evidence prediction. For pre- dicted evidence, rewriting raises mean F1 from 0.389± 0.006 to 0.398± 0.001, primarily through improved precision. C Implementation Details Stage 1. The 46-class nnU-Net-v2-style U-Mamba2 segmenter uses 128× 256× 256 patches and a 1,500-epoch schedule. In ad- dition to segmentation masks, it exposes 1,408-dimensional pooled encoder tokens for individual teeth and anatomical structures. Stage 2. The fixed tooth configuration performs label-wise sam- pling at the tooth-instance level. For every class and epoch, all supervised positives are retained and eligible negatives are sam- pled without replacement. The target negative-to-positive ratio is 0.5≤ n − /n + < 2 when enough negatives are available; otherwise all available negatives are retained. Sampling determines the active label-instance pairs, while their explicit or pu-weak reliability weights remain in the loss. The PU-weak weights are predefined reliability weights under the guidance of the clinical routine, not learned or tuned on validation data. Unmentioned negatives receive class-dependent weights reflecting how reliably non-mention im- plies absence. For example, if a mask-present tooth is not reported as missing, it is highly likely to be present, so missing receives a high weak-negative weight. In contrast, non-mention of caries is less reliable because subtle caries may be overlooked or selectively omitted, so caries receives a lower weight. For positions with a Stage 1 tooth mask, the mask center and principal axes define the oriented crop. Mask-absent positions re- ceive only a fixed crop centered by interpolation from teeth in the same quadrant and no tooth-branch supervision; their missing status is inherited directly from the Stage 1 mask during branch merging. For a classifiable position with tooth feature u, the missing proba- bility m = σ(h m (u)) conditions the remaining attributes through [u; m detach φ(u)], allowing missingness to coexist with residual or prosthetic findings without propagating attribute loss into the missing head. Each branch is trained independently. The main con- Schema fieldMeaningTotal n + 38_close_to_left_IACTooth 38 close to the left IAC168 48_close_to_right_IACTooth 48 close to the right IAC168 any_tooth_close_to_IACAny reported tooth-IAC proximity219 lesion_close_to_IACLesion close to an IAC14 left_IAC_predominantly_lingualPredominantly lingual course of the left IAC236 left_IAC_predominantly_buccalPredominantly buccal course of the left IAC60 right_IAC_predominantly_lingualPredominantly lingual course of the right IAC243 right_IAC_predominantly_buccalPredominantly buccal course of the right IAC61 Table 7: Complete relation prediction schema and positive sample counts. IAC denotes the inferior alveolar canal. Residual categoryCases Image orientation, laterality discrepancy, or acquisi- tion note 3 Low-frequency dental classification (Kennedy class)1 Explicit complete-dentition or complete-arch state- ment 2 Recurring omission of common dental findings0 Table 8: Substantive report content not fully represented by retained evidence spans. Counts denote affected evaluation cases; categories may overlap within a case. figuration uses a scratch global branch, a hybrid relation branch, and the gated hybrid tooth branch with the metal-sensitive channel and pu-weak supervision. Table 9 specifies the actual tensors and encoders used by the main configuration. Dimensions and physical extents follow (z,y,x). The global view covers the complete available scan and is resized to a fixed grid, so its effective spacing varies with the acquired field of view. For each tooth, PCA is applied to the 3D coordinates of its segmentation mask to define a tooth-centered local coordinate system; PCA is used only to estimate crop orientation, not for di- mensionality reduction. An oriented bounding box enclosing the complete tooth is then expanded by 6 m in all directions and resized to the fixed network input size. Because tooth dimensions vary across cases, this operation produces tooth-dependent effective voxel spacing. Relation crops instead use a fixed 80× 96× 64-m field of view; resizing it to 80 × 128 × 128 gives an effective resolution of 1.0× 0.75× 0.5 m. Stage 3. Before rendering, the selected record is mapped through the deterministic dental-logic projection e E = Π dent ( b E). A tooth- specific IAC-proximity tuple for position 38 or 48 is omitted when that position is marked missing, except when impaction or root- remnant evidence at the same position indicates residual dental structure. When predominantly lingual and predominantly buccal canal-course tuples are both selected on one side, the tuple with the higher predicted probability is retained. The projection uses only Stage 2 tuples and confidence values, and it never introduces a positive finding absent from b E. GPT-4o rewrites 561 training drafts constructed from consistency-projected predicted evidence. Qwen2.5-7B is adapted with LoRA rank 16 for three epochs using completion-only loss, and inference uses greedy decoding. The rewriter receives only the deterministic draft. D Rewrite Content-Preservation Audit An author manually reviewed 30 evaluation cases selected by sim- ple random sampling from the primary EviDent-CBCT run. For each case, the deterministic template and Qwen rewrite were com- pared proposition by proposition for additions, omissions, polarity changes, FDI reassignment, anatomical-side reassignment, finding- type substitutions, and redundant statements. The rewrite preserved all template propositions in 28 of 30 cases (93.3%). Table 10 sum- marizes the observed deviations. Both substantive departures moved the output toward the refer- ence report rather than introducing an unsupported reference-level finding. In case A014, the rewrite omitted a predicted prosthetic crown at tooth 24 and redundantly repeated endodontic treatment at tooth 36; the reference described a coronal restoration rather than a crown at tooth 24. In case A067, the rewrite changed a predicted coronal restoration at tooth 43 to a prosthetic crown, matching the reference. These observations indicate that rewriting is not strictly lossless and can occasionally regularize imperfect predicted evi- dence, consistent with its modest mean F1 improvement. The re- tained deterministic template provides an auditable fallback and exposes such changes directly. E Controlled Direct Generation Suite Controlled interfaces and supervision We instantiate the direct generators with general-purpose Qwen3- 8B (Yang et al. 2025) and clinically aligned Llama3-Med42-8B (Christophe et al. 2024). Both use LoRA rank 16, completion-only loss for 12 epochs, deterministic decoding, and the training and evaluation split defined in the main paper. Two targets are used throughout: raw denotes the observed report, and canonical de- notes the report obtained from reference-derived evidence through deterministic rendering and the frozen rewriter. Direct 1 projects 72 frozen Stage 1 structure tokens into the lan- guage model. Direct 2 recomputes those tokens from the fixed-size CBCT input and updates encoder levels 2-4. Direct 3 uses the same global, relation, and tooth views and branch architectures as Stage 2 and initializes them from the corresponding primary Stage 2 check- points. During direct SFT, these visual modules and their feature adapter remain trainable, and their continuous output is supplied directly to the language decoder without a discrete evidence inter- face. Direct 4 retains the complete Direct 3 pathway and adds evidence prediction heads. Its training objective is L direct+aux =L report + λ 3 (L global +L relation +L tooth ), (5) with λ = 1. Global and relation terms are masked BCE losses; the tooth term uses the same PU-weighted reliability values as the Stage 2 tooth labels. Ground-truth evidence is used only to form this training loss and is not supplied at inference. Direct 5 retains the Direct 3 continuous-feature interface but freezes all Stage 2 branch parameters at their evidence-trained checkpoints. Only the feature adapter and language-model LoRA BranchView and tensor sizeResolution/channelsEncoder and fusion GlobalWhole available scan;96×192×192 Case-dependent effective spac- ing; 1 normalized CT + 5 se- mantic channels Seven-stage3Dresidualencoderwithwidths [32, 64, 128, 256, 320, 320, 320], block counts [1, 3, 4, 6, 6, 6, 6], and strides [1, 2, 2, 2, 2, 2, 1]; terminal 320-channel MambaLayer2, global average pooling, and a 12-label MLP. This branch is trained scratch-only. ToothOne PCA-aligned crop per FDI posi- tion; 64× 64× 96, with the tooth long axis mapped to x Tooth-dependent spacing after adding 6 m on each crop axis; 1 anatomical CT + 1 metal-sensitive CT + 8 seman- tic channels SmallCropEncoder: three 3× 3× 3 convolutions with 12, 24, and 48 channels, instance normalization and LeakyReLU, pooling after the first two blocks, global pooling, and a 192-D projection. The crop feature is gated with a 192-D projection of the 1,408-D Stage 1 tooth token and a 16-D ten-region embedding, yielding a 256-D feature for label-specific heads; the missing head conditions the remaining attributes. RelationPaired left/right posterior-mandibular crops; each 80× 128× 128 from an 80× 96× 64-m FOV 1.0× 0.75× 0.5 m; 1 nor- malized CT + 5 semantic chan- nels The same 12-24-48 SmallCropEncoder produces a 192-D feature per side. Left/right features and 16-D side embeddings are combined with projected Stage 1 relation tokens by gated fusion into 384 dimensions, followed by an eight-label MLP. Table 9: Stage 2 input geometry and encoder architecture in the primary configuration. CT and semantic-mask channels are concatenated before the branch encoder. Audit outcomeCases All template propositions preserved 28 Any substantive content change2 Omission1 Finding-type substitution1 Polarity reversal0 FDI reassignment0 Anatomical-side reassignment0 Redundant repetition1 Table 10: Author audit of content changes introduced by Qwen rewriting. Categories can overlap within a case. parameters are optimized. This control separates the effect of pre- serving the Stage 2 representation from the effect of exposing a thresholded, discrete evidence record to the report generator. Run-level results The main paper reports logical precision, recall, F1, BLEU-4, and METEOR for the complete comparison. Table 11 provides logical- F1 for each model run. Canonical targets improve all ten decoder- interface pairs. Med42 improves the strongest direct result from .331 with Qwen3 to .371. Direct 5 reaches .327 with Qwen3 and .360 with Med42, and does not improve on trainable Direct 3 with either decoder. Case-level paired bootstrap We compare the final consistency-projected, locally rewritten EviDent-CBCT reports with the strongest direct configuration by unrounded mean logical-F1, Med42 Direct 2 with canonical su- pervision. For each system, case-level RadFact-Lite-Dental logical precision and recall are first averaged across its three model runs. Using random seed 2027, we jointly resample the matched case indices with replacement 100,000 times. Within every replicate, corpus-level precision and recall are averaged over the sampled cases and logical-F1 is recomputed as their harmonic mean. The percentile 95% interval is taken from the resulting paired differ- ences, F 1 EviDent − F 1 Direct . Using the final consistency-projected outputs and unrounded case-averaged values, the direct and EviDent-CBCT aggregate logical-F1 scores are .3706 and .4023, respectively. Their paired difference is .0317, with a 95% bootstrap confidence interval of [.0047, .0586]. This interval quantifies case-sampling uncertainty after averaging model runs; it is not a confidence interval over training seeds. The comparison between Direct 3 and Direct 4 separates auxil- iary evidence supervision from an evidence bottleneck. Adding the auxiliary objective changes Qwen3 logical-F1 from .236 to .244 for raw targets and from .331 to .331 for canonical targets; the corresponding Med42 changes are .280 to .271 and .362 to .365. Freezing the Stage 2 encoders in Direct 5 yields .234 and .327 with Qwen3, and .262 and .360 with Med42. Neither auxiliary supervi- sion nor preservation of continuous Stage 2 features reproduces the .402 obtained when the consistency-projected discrete record is the required report input. Sensitivity to multiple reports per volume The fixed-reference evaluation selects the report with the lexico- graphically smallest identifier for each volume. As an additional robustness analysis, we evaluate each candidate against every re- port associated with its volume; 39 of the 63 validation volumes have two reports and 24 have one (102 reports in total). If P ij and R ij are the logical precision and recall for candidate i against its j-th reference, we first compute ̄ P i = 1 m i m i X j=1 P ij , ̄ R i = 1 m i m i X j=1 R ij ,(6) then average ̄ P i and ̄ R i over the 63 volumes and compute their harmonic mean. Thus, every volume retains equal weight regardless of its number of reports. Under mean-over-references evaluation, both EviDent-CBCT re- alizations round to .374 logical-F1, compared with .333 for the strongest direct generator. For the final locally rewritten output, the unrounded paired advantage over direct generation is .0402, with a 95% bootstrap confidence interval of [.0170, .0637] using the same 100,000 volume-level resamples. The rewrite–template difference is negligible (−.0004, 95% CI [−.0074, .0065]), indicating that the advantage over direct generation persists independently of the local realizer and reference selection. F Global and Relation Fusion Sweep The main paper fixes the global branch to scratch and the relation branch to hybrid (g_scratch_r_hybrid). All rows share the FamilyVisual interfaceAuxiliary evidence TargetRun 1 Run 2 Run 3 Direct Qwen3-8B Direct 1: frozen Stage 1NoRaw.235 .242 .209 Direct 1: frozen Stage 1NoCanonical.325 .258 .297 Direct 2: fine-tuned Stage 1NoRaw.244 .265 .251 Direct 2: fine-tuned Stage 1NoCanonical.283 .355 .327 Direct 3: continuous Stage 2 featuresNoRaw.240 .258 .210 Direct 3: continuous Stage 2 featuresNoCanonical.328 .330 .337 Direct 4: Direct 3 + auxiliary evidence YesRaw.239 .214 .281 Direct 4: Direct 3 + auxiliary evidence YesCanonical.341 .315 .338 Direct 5: frozen Stage 2 encoderNoRaw.240 .239 .223 Direct 5: frozen Stage 2 encoderNoCanonical.326 .325 .330 Direct Med42-8B Direct 1: frozen Stage 1NoRaw.249 .240 .241 Direct 1: frozen Stage 1NoCanonical.341 .358 .334 Direct 2: fine-tuned Stage 1NoRaw.275 .304 .256 Direct 2: fine-tuned Stage 1NoCanonical.375 .351 .386 Direct 3: continuous Stage 2 featuresNoRaw.283 .260 .296 Direct 3: continuous Stage 2 featuresNoCanonical.358 .360 .370 Direct 4: Direct 3 + auxiliary evidence YesRaw.270 .271 .274 Direct 4: Direct 3 + auxiliary evidence YesCanonical.358 .364 .372 Direct 5: frozen Stage 2 encoderNoRaw.266 .269 .251 Direct 5: frozen Stage 2 encoderNoCanonical.363 .345 .371 EviDent-CBCTConsistency-projected evidence record –Template.401 .393 .401 Consistency-projected evidence record –Local rewrite .405 .399 .403 Table 11: Run-level RadFact-Lite-Dental logical-F1 for the controlled report-generation comparison. Reference protocolCandidate outputLogical PLogical RLogical F1 Fixed lexicographicMed42 Direct 2, canonical (direct best) .388± .019 .354± .017 .371± .018 EviDent-CBCT template.372± .002 .428± .008 .398± .004 EviDent-CBCT local rewrite.384± .002 .423± .004 .402± .003 Mean over all references Med42 Direct 2, canonical (direct best) .361± .013 .310± .013 .333± .013 EviDent-CBCT template.369± .007 .379± .012 .374± .009 EviDent-CBCT local rewrite.374± .003 .373± .011 .374± .006 Table 12: Sensitivity of the final comparison to the reference protocol, reported as mean±std over three model runs. Under mean-over-references evaluation, logical precision and recall are averaged within each volume before corpus aggregation. Global Relation Logical P Logical R Logical F1 hybrid hybrid.653.692.672 hybrid scratch.651.691.670 hybridvisual.651.695.672 scratch hybrid.652.696.673 scratch visual.650.699.674 visualhybrid.655.702.678 Table 13: Global/relation interface sweep using merged Stage 2 exact-tuple agreement. The narrow range does not support a claim that one fusion choice is intrinsically supe- rior. tooth-base branch. Table 13 reports a single-run architecture sweep using the LLM-free merged Stage 2 logical set metric. The observed logical-F1 range is 0.670-0.678. We retain g_scratch_r_hybrid as the predefined base for the tooth ab- lations and do not treat this narrow sweep as evidence for a general fusion mechanism. G Evaluation Semantics Merged Stage 2 logical set-F1 For case k, let C k be the set of predicted positive tuples and R k the positive tuples derived from its reference report, pooled across tooth, global, and relation branches within the evaluation mask. Exact structured matching gives P k = |C k ∩ R k | |C k | , R rec k = |C k ∩ R k | |R k | .(7) Zero denominators contribute zero, matching the RadFact-Lite ag- gregation convention. The reported logical precision and recall are the means of P k and R rec k over cases; logical set-F1 is the harmonic mean of those two corpus-level quantities. This is an LLM-free structured-agreement metric. It is not branch macro-F1 and is not the report-level RadFact score. RadFact-Lite-Dental RadFact-Lite-Dental is a dentist-guided evaluation extension built from RadFact-Lite v0.1.0 (Lumetti 2026), a lightweight implemen- tation of the text-only RadFact procedure (Bannur et al. 2024). It retains RadFact’s bidirectional proposition-entailment protocol and corpus aggregation while adapting proposition construction and en- tailment assessment to dento-maxillofacial reports. GPT-5.6-Luna first parses each report into complete, image-verifiable dental propo- sitions. The parser atomizes tooth lists but preserves qualifiers such as laterality, severity, extent, uncertainty, visibility, treatment status, and anatomical relations; bridge spans remain relational proposi- tions rather than independent tooth labels. Before LLM assessment, conservative normalization maps equivalent dotted and undotted FDI notation to a common form, removes semantic duplicates, and resolves exact or closed-form positive entailments. The deterministic layer never rejects an unre- solved comparison. Remaining hypotheses are assessed jointly for each case and direction with dental-specific instructions that require agreement in tooth position, anatomy, side, finding type, polarity, and clinically meaningful qualifiers. Parsing and entailment outputs are cached for identical normalized inputs, and this final RadFact- Lite-Dental configuration is used for all report-level evaluations. For report-level evaluation, the report with the lexicographically smallest identifier was fixed as the reference for each volume and used across all systems and evaluator runs. Official ODIN hidden-test RadFact uses a different evaluation protocol and is therefore reported separately without direct numer- ical comparison to RadFact-Lite-Dental. Author audit of RadFact-Lite-Dental. We conducted an author audit of 30 candidate-reference report pairs, sampling ten cases from each tertile of total proposition count. For each pair, the complete reports, proposition boundaries, and every bidirectional entailment decision were reviewed to establish author-audited gold scores. Table 14 compares these scores with the original RadFact- Lite v0.1.0 implementation using GPT-4o without dental FDI nor- malization and with the final RadFact-Lite-Dental evaluator. Un- der the same corpus-level aggregation, RadFact-Lite-Dental differs from the author-audited logical-F1 by −.009, with a paired case- bootstrap 95% confidence interval of [−.030,.012]. The original evaluator nearly reproduces the gold cohort-level F1 (.410 versus .410), but this agreement does not extend to indi- vidual cases: its case-level MAE is .091, Pearson r = .587, and maximum absolute error is .307. Its 19 overestimates and 11 under- estimates largely cancel in the aggregate. The final dental evaluator reduces MAE by 52.7% to .043, increases Pearson correlation to .881, and reduces the maximum absolute error to .140. Its Spear- man correlation is ρ = .861, and 20 of 30 cases fall within .05 F1 and 26 within .10, compared with 12 and 19 cases for the origi- nal evaluator. Across 1,375 fixed-proposition entailment decisions, the final evaluator reaches .949 accuracy, .941 balanced accuracy, .931 entailment F1, and Cohen’s κ = .891 against the author audit. Thus, its small aggregate offset accompanies substantially closer case-level agreement and more faithful case ordering. H Class-wise Stage 2 Branch Analysis The three supervision policies alter only the tooth branch; the crop geometry, gated hybrid encoder, metal-sensitive channel, global branch, and relation branch are otherwise fixed. We therefore com- pare pu-weak, explicit, and hard directly at the tooth inter- face. For a tooth category, a tuple is defined by its FDI position and finding label. Table 15 reports exact-tuple precision, recall, and F1 on the strict evaluation mask, with unmentioned eligible tooth attributes treated as negatives. The branch-level ordering matches the aggregate tooth-branch ablation in the main paper:pu-weak gives the highest tooth logical set-F1, followed by hard and then explicit. Its advantage is not uniform across every rare label, but it obtains the highest class F1 for six of the 11 categories, including missing teeth, crowns, endodontic treatment, prosthesis-related findings, impaction/disodontiasis, and caries. The explicit model often retains recall while losing pre- cision under the strict mask, most visibly for restorations (P=.161, R=.943) and caries (P=.014, R=.667), showing that ignored non- mentions leave many unsupported positives unconstrained. Full- weight negatives recover much of this precision, while reduced- reliability negatives provide a better aggregate balance. Root frac- ture, caries, and root remnant have only 1, 10, and 12 positive vali- dation instances, respectively; their policy-specific ordering should therefore be interpreted as descriptive rather than as a stable rank- ing. Table 16 completes the branch view for the fixed base system. Known is the number of eligible supervised entries and Positive is the number of reference-derived positives within those entries. The global branch is strongest for common coverage and atro- phy statements, reaching .884 branch logical set-F1. Osteosclerotic lesions remain precision-limited, while the perfect scores for pe- riodontitis and poor-quality scans are supported by no observed negatives and only 19 and two positives, respectively. In the rela- tion branch, tooth-IAC proximity and lingual-course labels are re- covered reliably on their eligible entries, whereas the less frequent buccal-course categories remain precision-limited. The lower re- lation logical set-F1 (.537) reflects case-level agreement over the complete predicted relation set and is therefore more demanding than the high pooled F1 of several individual, positive-heavy cate- gories. I Qualitative Report Comparisons Figures 3–6 compare representative EviDent-CBCT predictions with those of the best direct model. In each composite, the upper-left and upper-right panels show the respective predictions, while the lower panels repeat the reference report with annotations relative to the prediction above. Dark green denotes a supported match, light green a partially correct or underspecified statement, blue an un- supported or contradictory prediction, and red a reference fact not recovered by the corresponding prediction. Uncolored text provides linguistic context. EvaluationLogical P Logical R Logical F1 F1 MAE Pearson r Author-audited gold.364.468.410– RadFact-Lite v0.1.0.371.458.410.091.587 RadFact-Lite-Dental.366.444.401.043.881 Table 14: Agreement with author-audited gold scores on 30 stratified report pairs. MAE and Pearson correlation are computed over case-level logical-F1. pu-weak (base) explicit hard FindingLogical PLogical RLogical F1Logical PLogical RLogical F1Logical PLogical RLogical F1 Missing tooth.412± .021 .622± .087 .494± .025 .346± .027 .524± .075 .413± .016 .368± .024 .694± .084 .478± .002 Implant.325± .013 .617± .021 .426± .014 .432± .092 .543± .057 .473± .046 .338± .037 .580± .057 .426± .035 Crown.452± .013 .646± .065 .530± .012 .233± .101 .646± .053 .331± .111 .427± .031 .663± .079 .516± .005 Restoration.360± .018 .665± .026 .466± .011 .161± .002 .943± .010 .275± .004 .424± .030 .542± .048 .474± .016 Endodontic treatment .413± .036 .595± .097 .482± .010 .385± .083 .356± .035 .367± .049 .382± .042 .595± .043 .464± .035 Prosthesis-related.604± .024 .789± .029 .683± .005 .611± .167 .513± .042 .549± .065 .649± .042 .693± .054 .668± .007 Impacted/disodontiasis .750± .035 .920± .040 .826± .037 .685± .071 .853± .023 .759± .052 .756± .046 .893± .046 .817± .009 Caries.149± .160 .467± .321 .121± .029 .014± .001 .667± .153 .027± .001 .051± .019 .433± .115 .089± .026 Root remnant.095± .064 .111± .048 .090± .028 .104± .032 .361± .096 .159± .043 .045± .016 .306± .048 .078± .025 Root fracture.002± .002 .667± .577 .004± .004 .004± .001 1.000± .000 .009± .003 .001± .002 .333± .577 .002± .004 Lesion-related.105± .017 .444± .030 .170± .024 .045± .014 .427± .339 .074± .009 .127± .024 .325± .039 .180± .020 Branch logical set .434± .028 .538± .022 .480± .007 .405± .057 .328± .035 .359± .018 .408± .017 .548± .030 .467± .017 Table 15: Class-wise tooth-branch Stage 2 performance under the three non-mention policies, reported as mean±std over three model runs. The final row is the case-grouped branch logical set score over all tooth categories. BranchFindingKnown Positive Logical P Logical R Logical F1 GlobalPartial maxilla coverage5427.579.815.677 Maxilla excluded5721.654.810.723 Maxillary sinuses excluded3014.9331.000.966 Mandibular condyles excluded5958.9831.000.991 Periodontitis19191.0001.0001.000 Maxillary atrophy861.000.833.909 Mandibular atrophy3934.9441.000.971 Severe atrophy3212.750.750.750 Osteolytic lesion3317.778.824.800 Osteosclerotic lesion232.2861.000.444 Sinus pathology166.7501.000.857 Poor quality/artifact221.0001.0001.000 Global logical set–.890.878.884 Relation Tooth 38 close to left IAC1918.9471.000.973 Tooth 48 close to right IAC13131.0001.0001.000 Any tooth close to IAC20201.0001.0001.000 Lesion close to IAC111.0001.0001.000 Left IAC predominantly lingual2922.7861.000.880 Left IAC predominantly buccal276.263.833.400 Right IAC predominantly lingual2619.7601.000.864 Right IAC predominantly buccal255.364.800.500 Relation logical set–.512.563.537 Table 16: Class-wise Stage 2 performance of the base global and relation branches. Counts expose the support available for each category; branch logical-set rows are case-grouped and are not averages of the class rows. Reference ToothFairy4 Report The MANDIBLE is included in the scan volume. The following teeth are missing: 38, 36. Implant - prosthetic rehabilitation of tooth 36; post-and-core restorations and related prosthetic crowns on teeth 35, 45, 46; endodontic treatment on teeth 37, 46 with prosthetic reconstruction of the respective crowns; reconstruction of the crown of tooth 34. Conservative dental treatments also on teeth 47, 48. The UPPER MAXILLA is partially included in the scan volume. The following teeth are missing: 28, 25, 14, 15, 18. Implant - prosthetic rehabilitation of tooth 15; post-and-core restorations and respective prosthetic crowns on teeth 26, 24; crown reconstruction of tooth 27 (it is not well definable whether there is a fixed prosthetic bridge involving 27, 26, 24). Incongruous endodontic treatment on teeth 26, 24, 14, 16. The MAXILLARY SINUSES are partially included in the scan volume and are aerated. The MANDIBULAR CONDYLES are not included in the scan volume. Reference ToothFairy4 Report The MANDIBLE is included in the scan volume. The following teeth are missing: 38, 36. Implant - prosthetic rehabilitation of tooth 36; post-and-core restorations and related prosthetic crowns on teeth 35, 45, 46; endodontic treatment on teeth 37, 46 with prosthetic reconstruction of the respective crowns; reconstruction of the crown of tooth 34. Conservative dental treatments also on teeth 47, 48. The UPPER MAXILLA is partially included in the scan volume. The following teeth are missing: 28, 25, 14, 15, 18. Implant - prosthetic rehabilitation of tooth 15; post-and-core restorations and respective prosthetic crowns on teeth 26, 24; crown reconstruction of tooth 27 (it is not well definable whether there is a fixed prosthetic bridge involving 27, 26, 24). Incongruous endodontic treatment on teeth 26, 24, 14, 16. The MAXILLARY SINUSES are partially included in the scan volume and are aerated. The MANDIBULAR CONDYLES are not included in the scan volume. EviDent-CBCT Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally, and includes the maxilla with the caudal portions of the maxillary sinuses bilaterally. In the mandible, the right and left mandibular canals exhibit a regular course, predominantly lingual, and are in close proximity to teeth 48. There is notable atrophy of the alveolar processes in the edentulous regions of both the maxilla and mandible. Teeth 36, 37, and 38 are absent. Prosthetic rehabilitation, including bridge and pontic structures, involves teeth 34, 35, 36, 45, and 46, with prosthetic crowns present on teeth 35, 36, 37, 45, and 46. Coronal restorations are noted on teeth 34 and 47. Endodontic treatment has been performed on teeth 35, 37, 45, and 46. Periodontal disease affects the existing teeth. There are no definitive osteolytic or osteocondensing lesions observed. In the maxilla, teeth 15, 18, and 28 are absent. Prosthetic rehabilitation, including bridge and pontic structures, involves teeth 15, 16, 24, and 26, with prosthetic crowns on teeth 15, 16, 24, and 26. Coronal restorations are present on teeth 17 and 25. Endodontic treatment has been conducted on teeth 14, 16, 25, and 26. The maxillary sinuses are normally pneumatized. The mandibular condyles are excluded from the acquisition and are not visible. Best Direct Model Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processesbilaterally, with partial inclusion of the maxilla. In the mandibular region, both the right and left mandibular canals exhibit a regular course. There is noted atrophy of the alveolar processes in the edentulous areas. Teeth 36, 37, 38, 46, 47, and 48 are absent. Prosthetic rehabilitation, including bridge or pontic structures, involves teeth 35, 44, and 45, which have prosthetic crowns. Periodontal disease affects the remaining teeth, and areas of osteolysis are observed. Regarding the maxilla, it is partially included in the scan. Tooth 16 is absent, and tooth 23 has undergone endodontic treatment. Periapical lesions are present in teeth 12, 17, 23, 24, and 25. The maxillary sinuses appear normally pneumatized. The mandibular condyles are excluded from the acquisition and are not visible. . Figure 3: A treatment-rich case with partial maxillary coverage. EviDent-CBCT recovers broader reference-supported tooth- status, restorative, and endodontic findings than the best direct model, although both predictions contain unsupported or mismatched details. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact. Reference ToothFairy4 Report Moderate atrophy of the mandibular bone at the level of the molar and premolar sectors bilaterally. Absence of teeth 33-35-36-37- 38-44-45-46-47-48. Adequate endodontic treatments involving teeth 34 and 43. Teeth 34 and 43 have prosthetic crowns. Regular course of the mandibular canal. The maxillary bone is partially included in the scan volumes, with a moderate pattern of bilateral bone atrophy. Absence of teeth 12-14-16-17-18-24-25-26- 27-28. Presence of root remnants in positions 11-13-15 (containing endodontic filling material)-21-22. Presence of a prosthetic crown on tooth 23 with endodontic treatment. Mandibular condyles not included in the scan volumes. Reference ToothFairy4 Report Moderate atrophy of the mandibular bone at the level of the molar and premolar sectors bilaterally. Absence of teeth 33-35-36-37- 38-44-45-46-47-48. Adequate endodontic treatments involving teeth 34 and 43, Teeth 34 and 43 have prosthetic crowns. Regular course of the mandibular canal. The maxillary bone is partially included in the scan volumes, with a moderate pattern of bilateral bone atrophy. Absence of teeth 12-14-16-17-18-24-25- 26-27-28. Presence of root remnants in positions 11-13-15 (containing endodontic filling material)-21-22. Presence of a prosthetic crown on tooth 23 with endodontic treatment. Mandibular condyles not included in the scan volumes. EviDent-CBCT Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally, and includes the maxilla with the caudal portions of the maxillary sinuses bilaterally. In the mandible, both the right and left mandibular canals exhibit a regular course. There is notable atrophy of the alveolar processes in the edentulous regions. Teeth 33, 35, 36, 37, 38, 44, 45, 46, 47, and 48 are absent. Prosthetic rehabilitation, including bridge or pontic structures, involves teeth 34, 42, and 43, with a prosthetic crown on tooth 34 and a coronal restoration on tooth 43. Both teeth 34 and 43 have undergone endodontic treatment. Periodontal disease affects the remaining teeth, and areas of osteolysis are observed. In the maxilla, teeth 12, 13, 14, 15, 16, 17, 18, 22, 24, 25, 26, 27, and 28 are absent. Prosthetic rehabilitation, including bridges or pontics, involves teeth 21 and 22, with prosthetic crowns on these teeth. Endodontic treatment has been completed on teeth 14, 21, 22, 23, and 24. The maxillary sinuses are normally pneumatized. The mandibular condyles are excluded from the acquisition and are not visible. Best Direct Model Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally, with partial inclusion of the maxilla. In the mandibular region, both the right and left mandibular canals exhibit a regular course. There is severe atrophy of the alveolar processes in the edentulous areas of both the maxilla and mandible. Teeth 35, 36, 37, 38, 45, 46, 47, and 48 are absent. Prosthetic rehabilitation, including bridge or pontic structures, involves teeth 34, 44, and 54, with prosthetic crowns present on these teeth. Tooth 33 has undergone endodontic treatment. Caries are noted on tooth 44. Periodontal disease affects the remaining teeth, and areas of osteolysis are observed. Regarding the maxilla, it is partially included in the scan. Teeth 15, 16, 17, 18, 25, 26, 27, and 28 are absent. A root remnant is present for tooth 23. The maxillary sinuses are normally pneumatized. The mandibular condyles are excluded from the acquisition and are not visible. Figure 4: A heavily edentulous case with a complex treatment history. EviDent-CBCT recovers the complete set of mandibular missing teeth and correctly localizes several restorative and endodontic findings. The direct model omits or misclassifies more tooth-level findings, while both models remain challenged by maxillary root remnants. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact. Reference ToothFairy4 Report Mandible: tooth 3.8 and 4.8 are in complete bony impaction. The mandibular canal has a regular course and is in close relationship with tooth 3.8 and 4.8. The maxilla is not included in the acquisition and is not assessable. The mandibular condyles are not included in the scan. Reference ToothFairy4 Report Mandible: tooth 3.8 and 4.8 are in complete bony impaction. The mandibular canal has a regular courseand is in close relationship with tooth 3.8 and 4.8. The maxilla is not included in the acquisition and is not assessable. The mandibular condyles are not included in the scan. EviDent-CBCT Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally. In the mandibular region, the right and left mandibular canals exhibit a regular course, predominantly lingual. Notably, the mandibular canal is in close proximity to teeth 38 and 48. There is evidence of mandibular atrophy in the alveolar processes of the edentulous areas. Teeth 38 and 48 are impacted. Periodontal disease is present in the existing teeth, and areas of osteolysis are observed. The maxilla is not included in the scan volume. The mandibular condyles are excluded from the acquisition and are not visible. Best Direct Model Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally. In the mandibular region, the right and left mandibular canals exhibit a regular course. Notably, the mandibulargia is in close proximity to tooth 38. Tooth 48 is absent, while teeth 38 and 43 are impacted. Tooth 38 shows evidence of an periapical lesion, and areas of osteolysis are observed. The maxilla is not included in the scan volume. The mandibular condyles are excluded from the acquisition and are not visible. Figure 5: A mandibular-only case with bilateral complete bony impaction of teeth 38 and 48. EviDent-CBCT captures both impacted teeth and their relationships with the mandibular canal, but omits the complete-bony qualifier and introduces several unsupported findings. The direct model misclassifies tooth 48 as absent and adds unsupported tooth-level abnormalities. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact. Reference ToothFairy4 Report Mandible: The right mandibular canal is in close relationship with impacted tooth 48. The left mandibular canal is in close relationship with impacted tooth 38. No definite osteolytic or osteosclerotic lesions. Maxilla: The maxilla is partially included in the acquisition. Mandibular condyles: Excluded from the acquisition and not visible. Reference ToothFairy4 Report Mandible: The right mandibular canal is in close relationship with impacted tooth 48. The left mandibular canal is in close relationship with impacted tooth 38. No definite osteolytic or osteosclerotic lesions. Maxilla: The maxilla is partially included in the acquisition. Mandibular condyles: Excluded from the acquisition and not visible. EviDent-CBCT Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally, with partial inclusion of the maxilla. In the mandibular region, the right and left mandibular canals exhibit a regular course, predominantly lingual. Notably, the mandibular canal is in close proximity to teeth 38 and 48, both of which are impacted. Periodontal disease is present among the existing teeth. There are no definitive osteolytic or osteocondensing lesions observed. Regarding the maxilla, it is only partially captured in the scan. Teeth 11 and 21 are absent. The maxillary sinuses are not included in the acquisition. The mandibular condyles are excluded from the scan and are not visible. Best Direct Model Prediction The CT scan encompasses the mandible, excluding the condylar and coronoid processes bilaterally. In the mandibular region, the right and left mandibular canals exhibit a regular course. Teeth 38 and 48 are impacted. There are no definite osteolytic or osteocondensing lesions observed. The maxilla is not included in the scan volume. The mandibular condyles are excluded from the acquisition and are not visible. Figure 6: A case with bilateral canal–impacted-third-molar relationships and partial maxillary inclusion. EviDent-CBCT covers all core reference propositions, including both canal relationships, the negative osseous finding, and the acquisition coverage. The direct model identifies the impactions but omits their canal relationships and incorrectly reports that the maxilla is not included. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact.