Paper deep dive
CT-Flow: Orchestrating CT Interpretation Workflow with Model Context Protocol Servers
Yannian Gu, Xizhuo Zhang, Linjie Mu, Yongrui Yu, Zhongzhen Huang, Shaoting Zhang, Xiaofan Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/20/2026, 4:03:22 PM
Summary
The paper introduces CT-Flow, an agentic framework for 3D CT interpretation that utilizes the Model Context Protocol (MCP) to orchestrate specialized tools (navigation, measurement, segmentation, radiomics). It proposes CT-FlowBench, a benchmark for evaluating tool-use and multi-step reasoning in clinical workflows, demonstrating significant performance improvements over static LVLM baselines.
Entities (8)
Relation Signals (5)
CT-Flow → creates → CT-FlowBench
confidence 97% · We curate CT-FlowBench, the first large-scale instruction-tuning benchmark tailored for 3D CT tool-use
CT-Flow → uses → Model Context Protocol
confidence 95% · By leveraging the Model Context Protocol (MCP), CT-Flow shifts from closed-box inference to an open, tool-aware paradigm.
CT-Flow → outperforms → Baseline Models
confidence 92% · CT-Flow achieves state-of-the-art performance, surpassing baseline models by 41% in diagnostic accuracy
CT-FlowBench → derivedfrom → CT-RATE
confidence 90% · CT-FlowBench is built on the CT-RATE corpus inheriting a vast library of 3D chest CT scans
CT-Flow → employs → ReAct
confidence 88% · Following the ReAct paradigm, the orchestrator drives the clinical workflow by interleaving its internal ratiocination with active environmental probing.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent advances in Large Vision-Language Models (LVLMs) have shown strong potential for multi-modal radiological reasoning, particularly in tasks like diagnostic visual question answering (VQA) and radiology report generation. However, most existing approaches for 3D CT analysis largely rely on static, single-pass inference. In practice, clinical interpretation is a dynamic, tool-mediated workflow where radiologists iteratively review slices and use measurement, radiomics, and segmentation tools to refine findings. To bridge this gap, we propose CT-Flow, an agentic framework designed for interoperable volumetric interpretation. By leveraging the Model Context Protocol (MCP), CT-Flow shifts from closed-box inference to an open, tool-aware paradigm. We curate CT-FlowBench, the first large-scale instruction-tuning benchmark tailored for 3D CT tool-use and multi-step reasoning. Built upon this, CT-Flow functions as a clinical orchestrator capable of decomposing complex natural language queries into automated tool-use sequences. Experimental evaluations on CT-FlowBench and standard 3D VQA datasets demonstrate that CT-Flow achieves state-of-the-art performance, surpassing baseline models by 41% in diagnostic accuracy and achieving a 95% success rate in autonomous tool invocation. This work provides a scalable foundation for integrating autonomous, agentic intelligence into real-world clinical radiology.
Tags
Links
- Source: https://arxiv.org/abs/2603.00123v1
- Canonical: https://arxiv.org/abs/2603.00123v1
Trouble viewing inline? Open PDF directly →
Full Text
62,014 characters extracted from source content.
Expand or collapse full text
CT-Flow: Orchestrating CT Interpretation Workflow with Model Context Protocol Servers Yannian Gu 1∗ , Xizhuo Zhang 1∗ , Linjie Mu 1 , Yongrui Yu 1 , Zhongzhen Huang 1 , Shaoting Zhang 3† , Xiaofan Zhang 12† 1 Qing Yuan Research Institute, Shanghai Jiao Tong University, Shanghai, China 2 Shanghai Innovation Institute, Shanghai, China 3 Sensetime Research, Shanghai, China ∗ Equal contribution, † Corresponding authors Abstract Recent advances in Large Vision–Language Models (LVLMs) have shown strong potential for multi-modal radiological reasoning, partic- ularly in tasks like diagnostic visual question answering (VQA) and radiology report gener- ation. However, most existing approaches for 3D CT analysis largely rely on static, single- pass inference. In practice, clinical interpre- tation is a dynamic, tool-mediated workflow where radiologists iteratively review slices and use measurement, radiomics, and segmenta- tion tools to refine findings. To bridge this gap, we propose CT-Flow, an agentic frame- work designed for interoperable volumetric in- terpretation. By leveraging the Model Context Protocol (MCP), CT-Flow shifts from closed- box inference to an open, tool-aware paradigm. We curate CT-FlowBench, the first large-scale instruction-tuning benchmark tailored for 3D CT tool-use and multi-step reasoning. Built upon this, CT-Flow functions as a clinical or- chestrator capable of decomposing complex natural language queries into automated tool- use sequences. Experimental evaluations on CT-FlowBench and standard 3D VQA datasets demonstrate that CT-Flow achieves state-of-the- art performance, surpassing baseline models by 41% in diagnostic accuracy and achieving a 95% success rate in autonomous tool invoca- tion. This work provides a scalable founda- tion for integrating autonomous, agentic intelli- gence into real-world clinical radiology. 1 Introduction Computed Tomography (CT) is a cornerstone of modern diagnostic radiology (Al Zaabi et al., 2025). The subtle 3D radiographic patterns (e.g., small hemorrhages, early ischemic changes, faint ground- glass opacities) and measurement-dependent crite- ria (e.g., size thresholds, volumetric assessments, and attenuation values in Hounsfield units) often critically influence downstream triage and treat- ment decisions (Wang et al., 2024b; Mao et al., Figure 1: Comparison of 3D CT analysis paradigms. Left: Traditional End-to-End LVLMs rely on passive visual ingestion of 3D data, resulting in static textual outputs. Right: The proposed CT-Flow framework lever- ages the Model Context Protocol to transform the LLM into an active agent. It dynamically orchestrates special- ized tools to deliver precise, multi-modal diagnosis. 2025). The rapidly increasing volume and com- plexity of CT examinations have consequently fu- eled strong interest in automated and assistive in- terpretation systems. Recent advances in Large Vi- sion–Language Models (LVLMs) offer a promising avenue to alleviate physician workload (Li et al., 2023; Tu et al., 2024; Jiang et al., 2025). By learn- ing from large-scale paired imaging and clinical text data, these models have demonstrated notable capabilities in diagnostic visual question answering (VQA) (Zhang et al., 2023) and automated report generation (Hamamci et al., 2024a; Zhao et al., 2024a), highlighting their potential role in radio- graphic interpretation. Despite these advances, current LVLM-based approaches remain poorly aligned with real-world CT interpretation workflows. Most existing meth- arXiv:2603.00123v1 [cs.CV] 23 Feb 2026 ods treat CT volumes as static visual inputs and rely on end-to-end inference over pre-encoded rep- resentations. To handle 3D data, prior work ei- ther employs specialized 3D encoders (e.g., 3D Vision Transformers) to aggregate voxel-level fea- tures (Bai et al., 2024) or adapts 2D encoders to pro- cess serialized axial slices (Hamamci et al., 2024b). Although effective for capturing global context, these strategies inevitably introduce information bottlenecks (Wu et al., 2025a) that obscure fine- grained anatomical details and subtle radiographic cues that clinicians rely on for evidence-based de- cision making (Wu et al., 2025b). More fundamen- tally, clinicians rarely arrive at diagnoses through a single passive observation. Instead, CT inter- pretation is an inherently active and iterative pro- cess involving scrolling through slices, switching planes, probing voxel densities, measuring lesions, and invoking specialized tools for segmentation or radiomics analysis (Ritchie et al., 2025). Yet most existing LVLMs operate in a “read-only” mode, lacking the agency required for iterative verifica- tion and hypothesis refinement (Goswami et al., 2025; Friebe, 2025). To better reflect clinical real- ity, we argue that CT interpretation should be re- framed as an agentic problem rather than a purely perceptual one. Recent progress in large language models (LLMs) has demonstrated that equipping models with explicit tool access can substantially improve problem-solving capability and response reliabil- ity. The recently proposed Model Context Proto- col (MCP) (Anthropic, 2024) further advances this paradigm by providing a standardized interface that connects LLMs to external data sources and tools, reducing reliance on bespoke integrations. The rapid adoption of MCP in general software ecosys- tems suggests a promising direction for transform- ing LVLMs from static predictors into dynamic orchestrators capable of interfacing with clinical utilities in a scalable and standardized manner. In this work, we introduce CT-Flow, the first agentic framework that incorporates MCP to trans- form passive volumetric encoding into active, tool- mediated probing. CT-Flow integrates four MCP servers supporting navigation, measurement, seg- mentation, and radiomics analysis. Within this framework, 3D CT understanding is formulated not as a single perception task but as a sequential decision-making process. Given a diagnostic query, the model dynamically decomposes the task into a series of tool calls, such as multi-planar visualiza- tion, region-specific segmentation, and quantitative radiomics extraction, and grounds its reasoning in tool-verified evidence. This design mitigates the information bottlenecks inherent to end-to-end 3D LVLMs and aligns model behavior more closely with real clinical workflows. To support this paradigm, we introduce CT- FlowBench, a benchmark designed to evaluate agent trajectories in 3D CT clinical workflows. Un- like existing static VQA datasets that focus primar- ily on final-answer correctness, CT-FlowBench for- malizes CT interpretation as executable reasoning chains and provides supervision over intermediate decisions and tool usage. Building upon CT-Flow, we establish a human–AI collaborative pipeline in which experts and agents iteratively co-design and refine diagnostic queries along with their ground- truth trajectories. This process yields 2,000 sam- ples for training and 300 samples for evaluation. We conduct extensive experiments on estab- lished 3D CT VQA benchmarks as well as CT- FlowBench.LVLMs equipped with CT-Flow achieve substantial performance gains and outper- form specialized medical LVLMs, demonstrating the effectiveness of reframing CT understanding as agentic reasoning with tool use. Moreover, models fine-tuned on CT-FlowBench-train exhibit strong generalization and robustness, indicating a promis- ing new paradigm for developing clinically aligned AI systems for CT interpretation. We summarize our contributions as follows: •We propose CT-Flow, a novel agentic architec- ture that leverages the MCP to shift 3D medical analysis from passive encoding to active, tool- mediated probing, aligning model behavior with clinical workflows. •We introduce CT-FlowBench, the first bench- mark dedicated to training and evaluating med- ical agents on 3D CT workflows, providing a standardized testbed for agentic reasoning in ra- diology. •We demonstrate that CT-Flow yields substan- tial performance improvements while producing transparent, traceable, and clinically aligned rea- soning processes compared to end-to-end LVLM baselines. 2 Related Work 3D Volumetric Interpretation in Medical VLMs. Research on 3D medical vision-language modeling ToolSpace Assess and quantify the target structure in the left hemithorax that contains an air– fluid interface and is associated with adjacent volume loss. Thought: Open an orthogonal 3-plane MPR view centered at the volume midpoint using lung windows. Action: view_ortho Thought: Use an axial montage on lung windows across the thorax to find the slice range where the air–fluid cavity is maximal. Action: view_montage Thought: Run TotalSegmentatorfocusing on thoracic structures. Action: segment_total_anatomy Final Answer: Large-size class with fluid component in soft-tissue/water-range density Data Ingestion Navigation Advanced Analysis Detailed Observation View Slice View Ortho Montage Load Data Seg Detailed Observation Global Navigation Data Ingestio n Advanced Analysis MONAISimpleITKPyRadiomicsNibabel (i) DataConstruction (i) Architectures (i) Case 3D CT-Image Report Text Quantitative Analysis Spatial Mapping Diagnostic Inference [Question] Which of the following is the most appropriate next stein diagnosis? [Answer] C. Thyroid- stimulating hormone level Curated Candidates 3D-RAD VQA TrainingSet X X CT - Flow Benchmark CT - Flow Training Data Raw Data from CT-Rate Divided into Scenarios Load Mask Mip Hu Servers Infrastructures Figure 2: Overview of the CT-Flow framework. (i) Data Construction: The pipeline for raw data curation, trajectory synthesis, and the establishment of the CT-Flow benchmark. (i) Architectures: The system decouples the LLM orchestrator from the imaging environment via FASTMCP, bridging high-level servers with medical imaging infrastructures to provide a suite of atomic tools in the Tool Space. (i) Case Study: A demonstration of a Language- Action TrajectoryT. The orchestrator performs Active Probing by iteratively generating reasoning states (s t ), executing tool calls (a t ), and interpreting high-fidelity observations (o t ) to reach a grounded diagnostic answer. generally follows two paradigms for volumetric interpretation. Native 3D approaches directly en- code volumes with 3D backbones: RadFM (Wu et al., 2025a) unifies 2D and 3D images via 3D ViT. M3D (Bai et al., 2024) deals with volumet- ric data via 3D spatial pooling. However, they suffer from high computational cost and limited fine-grained detail retention (Ates et al., 2025). Serialization-based methods treat CT volumes as long slice sequences: OmniCT (Anonymous, 2025) aggregates cross-slice context via sequence model- ing but may lose spatial fidelity. Hulu-Med (Jiang et al., 2025) unifies 2D and 3D inputs within a transparent generalist medical VLM by decompos- ing 3D volumes into constituent slices. Vote-MI (Wang et al., 2024b) selects representative slices to reduce computation, at the risk of discarding diagnostically critical spatial cues. While these methods excel at capturing global patterns, their reliance on static, lossy encoding may hinder the re- tention of fine-grained details necessary for precise diagnosis (Zhong et al., 2025). Agentic Reasoning and Tool Orchestration. Leveraging an LLM as the reasoning backbone, an autonomous agent should interact with its envi- ronment to make decisions and take actions. Re- Act (Yao et al., 2022) serves as a general agent paradigm that combines reasoning and acting in a unified framework, which enables LLMs to con- duct reasoning for making plans and taking actions, such as tool invocation, and to incorporate infor- mation from the environment. To facilitate LLMs in invoking orchestrated tools for solving complex tasks, the MCP (Anthropic, 2024) provides an open and standardized protocol for managing and execut- ing tools. For example, AgentMaster (Liao et al., 2025) utilizes MCP to provide a unified interface for tool access, long-term memory, and context management for multi-modal information retrieval. MCP facilitates interactions between LLM reason- ing and specialized tool invocation. Medical Agents and Clinical Tool-use. Using Large Language Models (LLMs) as controllers for external tools has gained significant traction in medical diagnostics. Early systems like Chat- CAD (Wang et al., 2024a) and ChatCAD+ (Zhao et al., 2024b) pioneered the use of LLMs to inte- grate outputs from various computer-aided diagno- sis tools into interactive reports. Recent research, such as Med-Agents (Tang et al., 2024), explores multi-agent frameworks to facilitate collaborative clinical reasoning across different specialized do- mains. MedRAX (Fallahpour et al., 2025) demon- strates the potential of LLMs in performing strate- gic planning and complex reasoning for radiology- specific tasks. Other efforts focus on integrating visual grounding utilities (e.g., localized segmen- tation and detection) into LLM’s reasoning loop. These agentic frameworks typically aim to trans- form LLMs from simple text generators to clini- cal co-pilots capable of orchestrating diverse tools. However, many of them treat tool invocation as iso- lated or fragmented actions, which fails to capture the iterative nature of complex 3D analysis. 3 Methodology CT-Flow aims to transform LVLMs from static predictors into clinical workflow orchestrators, en- abling them to make diagnostic plans, invoke di- verse imaging tools on demand, and complete CT tasks through iterative verification. An overview of CT-Flow is illustrated in Fig. 2. 3.1Standardizing Clinical Interface via MCP To enable a model to reason across complex 3D vol- umes, we first define a standardized action space by abstracting heterogeneous imaging operations into a composable toolchain via the MCP. In practice, we consolidate high-frequency, clinically essential capabilities in CT workflows into four tool suites, spanning the full pipeline from data loading to pre- decision verification: •Data Ingestion: Ingests CT volumes and meta- data into a standardized, queryable 3D state for downstream tool execution. •Global Navigation: Enables fast whole-volume orientation and coarse anatomical localization to guide subsequent probing. • Detailed Observation: Retrieves targeted high- resolution views (slices or sub-volumes) to verify diagnostic hypotheses with local evidence. • Advanced Analysis: Provides quantitative and structured measurements (e.g., Hounsfield Units or segmentation) to support decisions. These tool suites collectively form the atomic action space. By encapsulating low-level image processing within the MCP framework, we pro- vide the necessary infrastructure for the model to perform iterative, goal-oriented probing, shifting the paradigm from static, single-step prediction to dynamic orchestration of clinical workflows. 3.2 Iterative Probing over ReAct Building upon the standardized action space de- fined above, we formalize the diagnostic process as a Reasoning-Acting Trajectory. Following the ReAct paradigm, the orchestrator drives the clinical workflow by interleaving its internal ratiocination with active environmental probing. For a given clinical queryQ, the system generates a sequential execution path: T =(s 0 , a 0 , o 0 ), (s 1 , a 1 , o 1 ), . . . , (s n , a n , o n ) wheres t denotes the reasoning state (the model’s thought process for interpreting findings and plan- ning the next step),a t ∈ Arepresents a specific action issued via the MCP interface, ando t is the observation (e.g., visual evidence or quantitative metrics) returned by the imaging environment. The significance of this trajectory-based formu- lation is that it transforms the diagnostic task into an active information-probing process. Instead of relying on a single-pass prediction, the orchestrator iteratively refines its understanding of the case. If a particular observationo t provides insufficient infor- mation to resolve the query, the model leverages its updated reasoning states t+1 to adjust its hypothe- sis and initiate further targeted probing. This iter- ative loop ensures that the final diagnostic answer Ais synthesized from the accumulated evidence collected throughout the trajectory, allowing the model to handle complex clinical cases that require multi-step examination. 4 Dataset Construction Here we describe the construction of CT- FlowBench, a benchmark that represents 3D chest CT cases as executable reasoning trajectories. Design Rationale.Clinical volumetric interpreta- tion is inherently interactive: radiologists navigate across slices, localize regions of interest, perform measurements, and iteratively integrate evidence. To reflect this workflow and enable step-level ver- ification, we construct CT-FlowBench with tool- mediated trajectories whose intermediate observa- tions are execution-retrievable from the raw vol- ume, rather than providing only final labels. 4.1 Data Source and Curation CT-FlowBench is built on the CT-RATE cor- pus (Hamamci et al., 2024b) inheriting a vast li- brary of 3D chest CT scans and expert-aligned radi- ology reports. Trajectory synthesis over the full CT- RATE corpus is computationally expensive, so we curate a subset that preserves high reasoning den- sity. Using heuristic filters, we prioritize cases with higher anatomical diversity, richer diagnostic con- tent, and stronger potential for quantitative assess- ment. This strategy concentrates CT-FlowBench on complex, multi-step reasoning trajectories that require cross-slice navigation and precise tool use, while excluding redundant cases that only support single-step reasoning or lack actionable clinical findings. Details of the curation procedure are pro- vided in Appendix. 4.2 Task Scenario Definitions. We define three complementary functional scenar- ios spanning fundamental perception to holistic reasoning to evaluate the model’s capabilities in discriminative precision and logical consistency. See the Appendix for details. Quantitative Analysis. This scenario assesses the model’s precision in identifying objective phys- ical properties (e.g., dimensions and attenuation values), by discriminating subtle numerical gradi- ents and technical constants from multiple-choice options while filtering out imaging noise. Spatial Mapping. This scenario focuses on the recognition of spatial topology and structural ad- jacency, requiring models to analyze relationships of sequential slices (e.g., contact, encasement, or displacement), and evaluates the ability to translate 3D visual details into precise spatial logic. Diagnostic Inference. This scenario requires models to correlate multi-dimensional findings and systemic observations, synthesize localized fea- tures into a unified global logic, and select the most coherent clinical inference. This assesses the ability to overcome local information bias through comprehensive holistic reasoning. 4.3 Trajectory Synthesis To bridge the gap between static task definitions and agentic navigation, we employ an Execution-in- the-loop Trajectory Synthesis strategy. Unlike tra- ditional VQA datasets that provide direct mappings from images to labels, CT-FlowBench requires the model to generate a sequence of interleaved thoughts, actions, and observations—collectively termed a “reasoning trajectory.” The construction of these trajectories follows an Execution-Feedback Refinement protocol. For each curated case, we utilize a teacher model (e.g., GPT-4o with specialized medical prompting) to ex- plore multiple potential reasoning paths. To ensure the quality and clinical validity of these paths, a trajectoryTis only included in the final benchmark if it satisfies the Procedural Consistency criterion: ∀(a i , o i )∈T , val(o i |V)∧ pred(T ) = y gt (1) In this formulation,a i ando i represent thei- th action (e.g., a specific slice navigation or ROI crop) and its corresponding observation. The func- tionis_validchecks whether the observationo i is physically grounded and retrievable from the raw volumetric data, whilepredict(T )ensures the en- tire logical chain terminates in the ground-truth diagnosis y gt . In total, we construct CT-FlowBench with 300 evaluation instances and approximately 2000 QA training instances for instruction tuning, where each training instance corresponds to one exe- cutable trajectory. The training set consists of two components: (1) trajectory-form instances from CT-Flow (based on a curated subset of CT-RATE) and (2) a subset of the 3D-RAD training data. We synthesize trajectories using multiple teacher models (Gemini-3-Pro-Preview/GPT-5.2/Claude- Sonnet-4.5) and apply execution-in-the-loop veri- fication with consistency-based filtering, retaining only instances whose intermediate observations are reproducible via tool execution on the raw volume and whose final answers match the annotations. 5 Experiments We evaluate CT-Flow on CT-FlowBench and 3D CT VQA dataset, 3D-RAD, to validate its diagnos- tic accuracy and tool-use autonomy. We conduct experiments to compare our agentic framework against state-of-the-art baselines and analyze the efficacy of its tool-mediated reasoning workflow. 5.1 Experimental Protocols Benchmarks.The evaluation of CT-Flow is con- ducted on the proposed CT-FlowBench and a re- fined subset of 3D-RAD VQA. To ensure a bal- anced and robust assessment, a stratified sampling of 200 instances per task is performed, with op- tion orders randomized for each query to neutralize positional bias (detailed in Appendix). Table 1: Performance comparison on 3D-RAD and CT-FlowBench. For 3D-RAD, the reported metrics represent the average performance across all sub-tasks. For CT-FlowBench, QA, AM, and D correspond to Quantitative Analysis, Spatial Mapping, and Diagnostic Inference, respectively. ModelsTool-use 3D-RADCT-FlowBench (Acc. %) BLEU-4 ROUGE-L B-Score LLM-Judge ACC(%)QAAMDDAvg. Leading-edge Models GPT-5.2✓22.0826.0685.5418.2563.5035.00 40.00 37.00 37.33 Gemini-3-Pro-Preview✓29.5935.5989.4626.3862.5945.00 43.00 44.0044.00 Claude-Sonnet-4.5✓20.1426.8185.718.8554.8344.00 44.00 43.00 43.67 Qwen3-VL-235B-A22B-Instruct✓20.5522.6285.449.5254.2130.00 36.00 36.00 34.00 GLM4.6-V✓23.2325.4781.5916.7152.0425.00 33.00 34.00 30.67 Specialized Medical VLMs M3D-LaMed-Llama-2-7B×12.3319.7086.9914.8824.1717.00 17.00 17.00 17.00 M3D-RAD×29.7637.39 91.3026.0058.0039.00 34.00 35.00 36.00 Hulu-Med-7B×12.7723.7186.7722.8461.2958.00 46.0037.00 47.00 Backbones Qwen2.5-VL-7B-Instruct✓18.3323.4677.697.5026.8314.00 22.00 21.00 19.00 Qwen3-VL-8B-Instruct✓20.8922.3180.0710.7749.0630.00 26.00 20.00 25.33 Supervised Fine-Tuned Models CT-Flow-7B✓36.6734.7389.3022.5061.3643.0052.0038.0044.33 CT-Flow-8B✓36.9637.4791.6523.6369.4642.0040.0047.0043.00 Baselines. We benchmark CT-Flow against a di- verse set of baselines, including: (i) Leading-edge models (GPT-5.2 (OpenAI, 2025), Gemini-3-Pro- Preview (DeepMind et al., 2023), Claude-Sonnet- 4.5, Qwen3-VL-235B-A22B-Instruct (Yang et al., 2025), and GLM4.6-V (Team et al., 2025)), (i) Ef- ficient VLM backbones (Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct) (Yang et al., 2024), and (i) Specialized medical models (M3D-LaMed- Llama-2-7B (Bai et al., 2024), M3D-RAD (Xin et al., 2025), and Hulu-Med-7B (Jiang et al., 2025)). Metrics We evaluate performance using accu- racy on multiple-choice questions. For open-ended tasks, we adopt (i) Automated linguistic metrics, including BLEU-4, ROUGE-L, and BERTScore, to measure semantic and structural alignment; and (i) LLM-as-a-Judge (average of Deepseek- V3 (DeepSeek-AI, 2025), Kimi-K2-Thinking, and GPT-OSS-120B (OpenAI, 2025)) protocol to pro- vide a high-level assessment. Implementations. All experiments are con- ducted on 4 NVIDIA H100 GPUs using the LLaMA-Factory (Zheng et al., 2024b) frame- work. For the fine-tuning stage, we perform full- parameter fine-tuning on Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct with a learning rate of1× 10 −5 , optimized using DeepSpeed ZeRO- 2 (Rasley et al., 2020) with a cosine learning rate decay schedule. For model deployment and eval- uation, we utilize SGLang (Zheng et al., 2024a) to serve the model and conduct inference via its official OpenAI-compatible APIs, ensuring high- throughput and efficient generation. 5.2 Comparison with SOTA Methods. The experimental results demonstrate the clear su- periority of domain-specific supervised fine-tuning (SFT) in elevating diagnostic precision and linguis- tic alignment. As shown in Table 1, the CT-Flow- 8B achieves SOTA performance on the 3D-RAD benchmark with an accuracy of 69.46%, a signifi- cant improvement over its base version and other general-purpose models. This performance gain is mirrored in the linguistic metrics, where SFT variants consistently outperform other models in BLEU-4, ROUGE-L, and BERTScore, indicating that specialized training allows the model to gen- erate reports that are not only factually accurate but also conform to professional clinical standards. The substantial +22.46% gain visualized in Figure 3 further confirms that while general reasoning is a strong foundation, domain-specific adaptation is es- sential for bridging the gap between general visual understanding and expert-level medical diagnosis. Simultaneously, our evaluation highlights the transformative impact of the CT-Flow agentic framework on general-purpose frontier models, while exposing the increased logical complexity of CT-FlowBench. Through tool-mediated reason- ing, general models such as GPT-5.2 and Gemini- 3-Pro-Preview effectively surpass specialized med- ical models like M3D-RAD (58.00%), achieving accuracies of 63.50% and 62.59% on 3D-RAD, respectively. However, the performance on CT- 40506070 CT-Flow(8B) GPT-5.2 Gemini-3-Pro-Preview Claude-4.5-Sonnet Qwen3-VL-235B- A22B-Instruct GLM4.6-V Qwen3-VL-8B-Instruct 47.0069.46 +22.46 55.1763.50 +8.33 55.5062.59 +7.09 42.0054.83 +12.83 44.1754.21 +10.04 46.6252.04 +5.42 47.0049.06 +2.06 Accuracy (%) 3D-RAD 20304050 28.6742.67 +14.00 33.0037.33 +4.33 33.6744.00 +10.33 38.6743.67 +5.00 29.6734.00 +4.33 25.3330.67 +5.34 28.6725.33 -3.34 CT-FlowBench w/ slice w/ CT-Flow M3D-RAD Figure 3: Comparative performance of various models using the CT-Flow framework vs. the slice-based baseline. FlowBench is notably lower across all models, with the highest average accuracy reaching only 44.33%. This discrepancy underlines the benchmark’s chal- lenge in evaluating multi-stage tool-use rather than simple visual recognition. Smaller models like Qwen3-8B even experience a slight performance regression on this benchmark, suggesting that the cognitive load of autonomous tool-calling and dif- ferential diagnosis requires either high-scale rea- soning capabilities or targeted fine-tuning to main- tain logical consistency throughout the workflow. 5.3 Analysis of Tool-mediated Reasoning. The experimental results illustrated in Figure 3 demonstrate a universal performance elevation across the 3D-RAD benchmark, confirming the effi- cacy of the CT-Flow framework in significantly en- hancing diagnostic accuracy. Most notably, the in- tegration of our agentic tool-use workflow enables the CT-Flow (SFT) model to achieve a state-of-the- art accuracy of 69.46%, which constitutes a sub- stantial +22.46% improvement over the slice-based baseline. This trend extends to general-purpose frontier models such as GPT-5.2 and Claude-Opus, which exhibit gains of +8.33% and +12.83% re- spectively. By leveraging tool-mediated reasoning, these general-domain models are able to signifi- cantly outperform specialized medical models like M3D-RAD (58.00%), proving that an agentic ap- proach can effectively bridge the domain expertise gap in medical imaging without the need for ex- haustive medical-specific pre-training. However, the results on the CT-FlowBench re- veal that the benefits of tool-mediated reasoning are closely tied to model scale and the inherent com- plexity of the reasoning task. While high-capacity models like Gemini-3-Pro and GPT-5.2 continue to show strong positive gains (+10.33% and +14.00% respectively), the smaller Qwen3-8B model ex- hibits a performance regression of -3.34%. This divergence suggests that the multi-step diagnostic logic required by the benchmark—encompassing Spatial Mapping, Quantitative Analysis, and Diag- nostic Inference—places a high cognitive demand on instruction-following and strategic planning that may exceed the zero-shot capabilities of smaller backbones. Despite this scaling effect, the consis- tent upward trajectory across the majority of mod- els confirms that decomposing complex queries into tool-mediated steps is a far more robust strat- egy for 3D medical analysis than traditional static slice feeding, provided the underlying model pos- sesses sufficient reasoning depth to manage the autonomous tool-calling workflow. 5.4 Tool-use Performance & Ablations Tool-use Performance Analysis. Table 2 pro- vides a comprehensive statistical breakdown of tool-calling behaviors across various models on the 3D-RAD and CT-FlowBench datasets. The results indicate a direct correlation between task complexity and the frequency of tool interactions. For instance, Claude-Sonnet-4.5 and Qwen3-VL- 235B exhibit a high average number of tool calls, particularly in the CT-FlowBench, where complex diagnostic workflows require multi-step reasoning and iterative data retrieval. Regarding reliability, GPT-5.2 sets the bench- mark for instruction following, maintaining near- zero tool-name errors and minimal argument hal- lucinations. This stability is crucial in clinical settings where precise parameter input is non- Models 3D-RADCT-FlowBench Calls Name Errors Args ErrorsCallsName Errors Args Errors GPT-5.24.1330.0060.0567.1930.0030.108 Gemini-3-Pro-Preview5.0880.0080.1566.6670.1000.451 Claude-Sonnet-4.55.9300.0020.0929.4900.0170.407 Qwen3-VL-235B-A22B-Instruct 6.0720.0850.1906.7190.2240.453 Qwen3-VL-8B-Instruct5.9630.7820.21111.1960.9690.385 GLM4.6-V4.0300.0270.0786.3140.0930.204 CT-Flow-7B4.0130.0070.0186.1700.0070.033 CT-Flow-8B4.2480.0070.0577.4800.0270.282 Table 2: Tool usage and error statistics on 3D-RAD and CT-FlowBench. Calls, Name Errors, and Args Errors denote the average number of tool calls, tool-name errors and tool-argument errors per case, respectively. negotiable. Conversely, smaller-scale models such as Qwen3-VL-8B-Instruct show significant degra- dation in performance; the high frequency of name errors (0.782) suggests that these models strug- gle to maintain the logical consistency required for long-chain tool manipulation in 3D medical spaces.These findings underscore that while vision-language alignment is necessary, the capa- bility to orchestrate professional tools is distinctly scaling property of larger, more sophisticated. Ablation Study on Tool Categories.To evaluate the structural integrity and necessity of our pro- posed toolset, we conducted a systematic ablation study by categorizing the tools into four functional tiers. In our framework, Data Ingestion serves as the foundational prerequisite for any task execution, while the other three categories are incrementally added or removed to observe success rates. The experimental results 4 demonstrate that each category plays an indispensable role in the clini- cal reasoning pipeline. The removal of Advanced Analysis tools leads to a failure in synthesizing quantitative clinical metrics, such as calculating flow velocities or volumetric ratios, which are es- sential for definitive diagnosis. Without Detailed Observation capabilities, the system’s sensitivity to micro-lesions and subtle anatomical anomalies drops sharply, proving that global context alone is insufficient for high-precision radiology. Fur- thermore, the absence of Global Navigation results in “spatial disorientation” within the model’s rea- soning process, as it loses the ability to efficiently index and transition between disparate 3D slices or sequences. Ultimately, the ablation results con- firm that our toolset is not merely a collection of independent functions but a coherent, hierarchical system. We conclude that all four tool categories QAAMDD 0 10 20 30 40 50 60 Accuracy (%) CT-Floww/o cls.2w/o cls.3w/o cls.4Format Errors 0.0 2.5 5.0 7.5 10.0 12.5 15.0 # Format Errors Figure 4: Performance impact of tool category ablation. Bars indicate accuracy (%) and the red line tracks for- mat errors. Removing specific tool classes (cls. 2-4) leads to decreased diagnostic accuracy and increased errors across all tasks, validating the necessity of the full hierarchical toolset. are essential and rationally designed for addressing the inherent challenges of 3D medical images. 6 Conclusion We present CT-Flow, a novel agentic framework that shifts the paradigm of 3D CT analysis from passive visual ingestion to protocol-driven Ac- tive Probing. By integrating the MCP and the trajectory-based CT-FlowBench, our approach en- ables an orchestrator to dynamically invoke spe- cialized clinical tools, effectively overcoming the information bottlenecks of traditional 3D-LVLMs. Empirical results demonstrate that CT-Flow not only achieves state-of-the-art performance across diagnostic benchmarks but also provides a trans- parent and traceable reasoning process that aligns with real-world radiological workflows, offering a promising path toward more interactive and reliable AI-assisted medicine. Limitations Despite the significant advancements made by CT- Flow in agentic clinical interpretation, this work has certain limitations. First, the current model relies primarily on supervised fine-tuning (SFT) on CT-FlowBench and has not yet integrated reinforce- ment learning (RL) to further optimize decision trajectories or minimize redundant tool invocations. Recognizing the potential of Reinforcement Learn- ing from Clinical Feedback (RLCF) in achieving expert-level diagnostic efficiency, we plan to in- corporate algorithms such as Proximal Policy Op- timization (PPO) or Direct Preference Optimiza- tion (DPO) in future iterations to refine reason- ing chains. Second, while this iterative multi-step paradigm provides in-depth analysis, its inference latency is inherently higher than that of traditional single-pass models, which may pose challenges in time-sensitive clinical scenarios such as acute stroke triage. Consequently, improving inference efficiency through parallel processing or trajectory pruning, while maintaining diagnostic precision and transparency, will be a key focus of our future research. Ethical Considerations Clinical Oversight and AgencyWhile CT-Flow transitions from passive inference to active, tool- mediated orchestration, it is designed strictly as a clinical decision support system (CDSS) rather than an autonomous diagnostic entity. The “Thought-Action Trajectories” generated by the model are intended to provide clinicians with a transparent reasoning path and tool-verified evi- dence. However, the risk of “action hallucinations” in which the agent might invoke incorrect tools or misinterpret tool outputs persists. Therefore, all automated tool sequences and final diagnostic suggestions must be reviewed and validated by cer- tified radiologists. We advocate for a “human-in- the-loop” deployment to ensure patient safety. Transparency and Accountability By leverag- ing the Model Context Protocol (MCP), CT-Flow provides an auditable trail of how a diagnostic con- clusion was reached. This traceability is a key ethical safeguard, allowing clinicians to inspect which slices were visualized and which radiomic features were extracted. Despite this, users should remain vigilant regarding the model’s performance across different patient demographics and imaging protocols. Data Privacy and Provenance CT-FlowBench is constructed using de-identified, publicly avail- able 3D CT datasets. Throughout the curation and instruction-tuning process, we have strictly adhered to ethical guidelines to ensure that no Protected Health Information (PHI) is included. All data processing workflows are compliant with standard medical data privacy regulations (e.g., HIPAA), and the framework is designed to be deployed in secure, local clinical environments to prevent unauthorized data exposure. References Adhari Al Zaabi, Rashid Alshibli, Abdullah AlAmri, Ibrahim AlRuheili, and Syaheerah Lebai Lutfi. 2025. Trends and trajectories in the rise of large language models in radiology: Scoping review. JMIR Medical Informatics, 13(1):e78041. Anonymous. 2025. OmniCT: Towards a unified slice- volume LVLM for comprehensive CT analysis. In Submitted to The Fourteenth International Confer- ence on Learning Representations. Under review. Anthropic. 2024. Introducing the model context proto- col. Accessed: 2024-11-25. Gorkem Can Ates, Yu Xin, Kuang Gong, and Wei Shao. 2025. Dcformer: Efficient 3d vision-language mod- eling with decomposed convolutions. arXiv preprint arXiv:2502.05091. Fan Bai, Yuxin Du, Tiejun Huang, Max Q-H Meng, and Bo Zhao. 2024. M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578. Google DeepMind and 1 others. 2023.Gemini: A family of highly capable multimodal models. https://arxiv.org/abs/2312.11805. Preprint, arXiv:2312.11805. DeepSeek-AI. 2025. Deepseek-v3.2: Pushing the fron- tier of open large language models. Adibvafa Fallahpour, Jun Ma, Alif Munim, Hong- wei Lyu, and Bo Wang. 2025.Medrax: Medi- cal reasoning agent for chest x-ray. arXiv preprint arXiv:2502.02673. Michael Friebe. 2025. Ai in radiology and interventions: a structured narrative review of workflow automation, accuracy, and efficiency gains of today and what’s coming. International Journal of Computer Assisted Radiology and Surgery, pages 1–10. Debanjan Goswami, Ronast Subedi, and Shayok Chakraborty. 2025. Medivlm: A vision language model for radiology report generation from medical images. In Findings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 10287– 10304. Ibrahim Ethem Hamamci, Sezgin Er, and Bjoern Menze. 2024a. Ct2rep: Automated radiology report genera- tion for 3d medical imaging. In International Con- ference on Medical Image Computing and Computer- Assisted Intervention, pages 476–486. Springer. Ibrahim Ethem Hamamci, Sezgin Er, Chenyu Wang, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Omer Faruk Durugol, Ben- jamin Hou, Suprosanna Shit, and 1 others. 2024b. Developing generalist foundation models from a mul- timodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834. Songtao Jiang, Yuan Wang, Sibo Song, Tianxiang Hu, Chenyi Zhou, Bin Pu, Yan Zhang, Zhibo Yang, Yang Feng, Joey Tianyi Zhou, and 1 others. 2025. Hulu- med: A transparent generalist model towards holis- tic medical vision-language understanding. arXiv preprint arXiv:2510.08668. Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Nau- mann, Hoifung Poon, and Jianfeng Gao. 2023. Llava- med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural In- formation Processing Systems, 36:28541–28564. Callie C Liao, Duoduo Liao, and Sai Surya Gadiraju. 2025. Agentmaster: A multi-agent conversational framework using a2a and mcp protocols for multi- modal information retrieval and analysis. In Proceed- ings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstra- tions, pages 52–72. Yuren Mao, Wenyi Xu, Yuyang Qin, and Yunjun Gao. 2025. Ct-agent: A multimodal-llm agent for 3d ct radiology question answering. arXiv preprint arXiv:2505.16229. OpenAI. 2025. Gpt-5 system card. Technical report, OpenAI. Accessed: 2026-01-02. OpenAI. 2025. gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925. Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. 2020. Deepspeed: System optimiza- tions enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD international conference on knowl- edge discovery & data mining, pages 3505–3506. Brandon Ritchie, Lesley Summerville, Max Sheng, Monica Choi, Sree Tirumani, and Nikhil Ramaiya. 2025. Impact of turnaround time in radiology: The good, the bad, and the ugly. Current Problems in Diagnostic Radiology. Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. 2024. Medagents: Large language models as collaborators for zero-shot medical rea- soning. In Findings of the Association for Computa- tional Linguistics: ACL 2024, pages 599–621. GLM Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, and 152 others. 2025. Glm-4.5: Agentic, reason- ing, and coding (arc) foundation models. Preprint, arXiv:2508.06471. Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaek- ermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, and 1 others. 2024. Towards generalist biomedical ai. Nejm Ai, 1(3):AIoa2300138. Sheng Wang, Zihao Zhao, Xi Ouyang, Tianming Liu, Qian Wang, and Dinggang Shen. 2024a. Interactive computer-aided diagnosis on medical image using large language models. Communications Engineer- ing, 3(1):133. Yuli Wang, Yuwei Dai, Craig Jones, Haris Sair, Jinglai Shen, Nicolas Loizou, Wen-Chi Hsu, Maliha Imami, Zhicheng Jiao, Paul Zhang, and 1 others. 2024b. En- hancing vision-language models for medical imaging: bridging the 3d gap with innovative slice selection. Advances in Neural Information Processing Systems, 37:99947–99964. Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Hui Hui, Yan- feng Wang, and Weidi Xie. 2025a. Towards gener- alist foundation model for radiology by leveraging web-scale 2d&3d medical data. Nature Communica- tions, 16(1):7866. Jing Wu, Yuli Wang, Zhusi Zhong, Weihua Liao, Na- talia Trayanova, Zhicheng Jiao, and Harrison X Bai. 2025b. Vision-language foundation model for 3d medical imaging. npj Artificial Intelligence, 1(1):17. Yu Xin, Gorkem Can Ates, Kuang Gong, and Wei Shao. 2025.Med3dvlm: An efficient vision-language model for 3d medical image analysis. arXiv preprint arXiv:2503.20047. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayi- heng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 oth- ers. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jian- hong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 22 oth- ers. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations. Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weix- iong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2023. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415. Weike Zhao, Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2024a. Ratescore: A metric for radiology report generation. arXiv preprint arXiv:2406.16845. Zihao Zhao, Sheng Wang, Jinchen Gu, Yitao Zhu, Lanzhuju Mei, Zixu Zhuang, Zhiming Cui, Qian Wang, and Dinggang Shen. 2024b. Chatcad+: To- ward a universal and reliable interactive cad us- ing llms. IEEE Transactions on Medical Imaging, 43(11):3755–3766. Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. 2024a. Sglang: Efficient execution of structured language model programs. Preprint, arXiv:2312.07104. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024b. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand. Association for Computa- tional Linguistics. Zhusi Zhong, Yuli Wang, Jing Wu, Wen-Chi Hsu, Vin Somasundaram, Lulu Bi, Shreyas Kulkarni, Zhuoqi Ma, Scott Collins, Grayson Baird, and 1 others. 2025. Vision-language model for report generation and out- come prediction in ct pulmonary angiogram. NPJ Digital Medicine, 8(1):432. A Appendix A.1 Provided Tools in CT-Flow To facilitate a streamlined and rigorous medical image analysis pipeline, CT-Flow provides a com- prehensive suite of specialized tools categorized into four functional modules. As summarized in Table 3, these modules are designed to mirror the natural clinical workflow—moving from raw data validation to holistic screening, followed by pre- cise local quantification, and finally culminating in AI-driven prognostic modeling. This modular architecture ensures both the reproducibility of the analysis and the scalability of the system for large- scale clinical studies. A.2 Details of Dataset Construction This appendix provides a comprehensive descrip- tion of the systematic pipeline developed to trans- form unstructured clinical CT reports into a high- precision, tool-calling benchmark. The construc- tion logic is designed to move beyond simple question-answering, focusing instead on the agent’s ability to execute complex, multi-step clinical rea- soning within a constrained action space. A.2.1 Multi-Agent Annotation and Taxonomic Categorization The process begins with the raw ingestion of clin- ical reports, which are processed by a specialized Medical Data Annotation Expert agent. The pri- mary objective of this stage is to filter and cat- egorize reports based on their potential to chal- lenge an AI agent’s reasoning capabilities through a weighted scoring system. The agent evaluates reports based on three pri- mary features: Numerical Presence (essential for quantitative precision tasks), Anatomical Diver- sity (requiring cross-regional spatial logic), and Clinical Uncertainty (necessitating autonomous decision-making). Based on these scores, reports are classified into three types: Type A (Atomic) for simple fact retrieval, Type B (Logical) for spa- tial relational tasks, and Type C (Autonomous) for complex reasoning. Reports containing explicit diagnostic uncertainty or follow-up recommenda- tions are prioritized as Type C to ensure the bench- mark includes cases that simulate a radiologist’s diagnostic synthesis. A.2.2High-Precision Ground Truth Synthesis Once a report is categorized, a Senior Radiologist Agent acts as a data architect to convert clinical Table 3: Classification and Functional Description of Medical Imaging Analysis Tools CategoryIncluded ToolsClinical Objective & Description I. Data Ingestion load_data,load_mask, inspect_metadata, inspect_mask_labels, search_anatomy_names, list_window_presets Data Integrity: Establishes the workspace and vali- dates metadata (modality, spacing, labels) to ensure subsequent analysis is performed on accurate objects. I. Global Observation view_montage,view_mip, view_minip, view_avgip Macro-Navigation: Provides a “bird’s-eye view” us- ing projection (MIP/MinIP) or array layouts to help users rapidly screen for high-density (nodules) or low- density (airways) abnormalities. I. Detailed Measurement view_slice ,view_ortho, measure_distance, measure_max_diameter, find_organ_center, extract_vessel_centerline, auto_crop_body, edit_geometry Spatial Localization: The most high-frequency mod- ule. Combines slice-by-slice exploration with 3D multi-planar reconstruction (MPR) and precise physi- cal quantification of lesion size and geometry. IV. Advanced Analytics segment_total_anatomy, analyze_hu_distribution, analyze_shape_properties, extract_radiomics_signature, visualize_radiomics_chart, analyze_lesion_texture Quantitative Biomarkers: Post-processing stage uti- lizing AI and statistics to transform visual data into structured radiomic features for prognostic modeling and malignancy assessment. findings into a structured JSON “Gold Standard”. This transformation is governed by the Hard Ev- idence Principle, which dictates that the agent must only extract findings supported by explicit measurements, specific anatomical descriptors, or clear Hounsfield Unit (HU) density values. Furthermore, this stage involves Tool-Logic Mapping.For every ground-truth fact, the radiologist agent defines the specific sequence of tools from the internal library required for verification.For instance, a finding re- garding a pulmonary nodule is programmat- ically linked tosegment_total_anatomyand measure_max_diameterfunctions. This ensures that the benchmark evaluates not only the final an- swer but also the validity of the agent’s execution path. A.2.3 Clinical Scenario Domain Configuration To ensure a broad evaluation across radiologic spe- cialties, each benchmark task is assigned to a spe- cific clinical scenario. This allows for granular analysis of where an AI agent may excel or fail. •Scenario A (Quantitative Precision): Fo- cuses on RECIST 1.1 compliance, absolute metric accuracy, and HU density analysis to test mathematical and volumetric precision. •Scenario B (Surgical Mapping): Focuses on the assessment of vascular encasement angles and tumor-vessel cleavage planes, testing the agent’s ability to provide preoperative spatial intelligence. • Scenario C (Diagnostic Synthesis): Focuses on systemic staging (TNM), enhancement pat- tern analysis, and complex pattern recognition across multiple organ systems. A.2.4 The “Blind Verification” Instruction Strategy A critical challenge in evaluating AI agents is pre- venting “textual cheating”, where the agent uses its internal linguistic knowledge to guess findings from the report text. To mitigate this, we imple- mented a Blind Verification strategy. During instruction generation, the agent is for- bidden from using explicit pathology labels, such as “the pancreatic tumor.” Instead, it must de- scribe the target using Spatial Landmark Ref- erencing (e.g., “the lesion 10m inferior to the Splenic Vein”). By removing explicit labels, the agent is forced to use visualization and segmen- tation tools to navigate the 3D volume and locate the target before it can perform any analysis. This ensures the benchmark measures true tool-calling and image-grounding capabilities. A.2.5 Action Space Constraints and SOP Validation The final stage of construction ensures that every generated task is solvable within the Action Space of the predefined CT-AI Tool Library. For every task, the pipeline generates a Standard Operat- ing Procedure (SOP) ground truth, representing the optimal sequence of tool calls a human expert would utilize to reach the conclusion. This SOP serves as a trajectory gold standard. By comparing the AI agent’s actual tool-calling sequence against this expert-defined SOP, we can evaluate the efficiency, medical logic, and safety of the agent’s diagnostic process, rather than relying solely on the accuracy of the final multiple-choice selection. A.3 Comparative Analysis: 2D Slice-based vs. CT-Flow Tool Processing The experimental results in Table 4 and Table 5 highlight a critical performance gap between stan- dard 2D slice-based vision processing and our pro- posed CT-Flow tool-enhanced framework. While baseline models process individual 2D slices with- out a global spatial understanding, the CT-Flow tool enables these models to integrate 3D spatial- contextual information. A.3.1 Bridging the 2D-to-3D Gap (Table 1) Analysis of the 3D Rad 1.2k dataset reveals that the lack of spatial context in the slice-based baseline severely limits diagnostic and descriptive perfor- mance: •Overcoming Information Fragmentation: Baseline models (Gemini-3-Pro, GPT-5.2, etc.) operating on individual slices struggle with Tasks 1–3 (Report Generation). For ex- ample, Gemini-3-Pro’s BLEU and Rouge scores more than double when utilizing the CT-Flow tool (BLEU:11.72→ 26.85; Rouge: 15.81→ 38.29). This suggests that while in- dividual slices contain local features, the tool provides the necessary spatial continuity to generate coherent clinical reports. •Diagnostic Accuracy through Spatial Syn- thesis: In Tasks 4–6, which require identi- fying volumetric features, the slice-based ap- proach shows significant limitations. The in- tegration of CT-Flow allows models to syn- thesize information across slices, leading to substantial accuracy gains. For Claude-Opus, the tool-enhanced version improves Task 5 ac- curacy by 23.19%, demonstrating that spatial relationships are essential for accurate medi- cal classification. •Efficiency vs. Specialization: Notably, the tool-enhanced general-purpose models signif- icantly outperform M3D, a model natively de- signed for 3D volumes. This indicates that the combination of high-level reasoning in LLMs and our spatial-contextual tool is more effec- tive than specialized architectures that process 3D data in a black-box manner. A.3.2 Scenario-Based Robustness (Table 2) Table 5 further illustrates how the tool-based ap- proach enhances model stability across different evaluation scenarios compared to the slice-only method: •Consistent Enhancement: Across all scenar- ios, the transition from 2D slice processing to CT-Flow tool integration leads to an upward trend. Gemini-3-Pro achieved an average im- provement of 10.33 points, reaching a peak of 45.00 in Scenario 1. This confirms that the tool provides a robust spatial prior that remains effective regardless of the specific scenario configuration. •Addressing Slice-Level Ambiguity: The baseline models often exhibit high variance between scenarios (e.g., Qwen3-VL-8B drop- ping from 31.00 to 22.00). The CT-Flow tool acts as a “spatial stabilizer”, helping models resolve ambiguities that occur when a lesion or anatomical structure is only partially visi- ble or appears differently across various 2D slices. •Fine-tuning vs.Tool Integration: In- terestingly, the Fine-tuned model—which was explicitly trained to handle these vol- umes—shows the highest Scenario 3 per- formance (46.00). However, the fact that Gemini-3-Pro + CT-Flow (44.00) reaches a similar level without scenario-specific train- ing underscores the power of providing the model with a contextual tool rather than just raw image slices. A.3.3 Summary of Findings The data provides strong empirical evidence that 2D slice-based processing is insufficient for com- plex 3D medical tasks. By utilizing the CT-Flow tool, models can leverage spatial-contextual infor- mation to achieve a more holistic understanding of the 1.2k dataset, resulting in performance that rivals or exceeds natively 3D-aware models and approaches the ceiling of fine-tuned systems. A.4 Case Study S You are an expert AI Medical Imaging Assistant, operating as a sophisticated ReAct(Reasoning and Acting) agent. Your mandate is to answer the user's question by actively interacting with medical imaging data using the provided tools. --- ### CORE PRINCIPLES 1. **Visual Evidence First**: Your thought process must explicitly describe what is visible in the tool output. If you do notdescribe it, you effectively did not see it. 2. **Clinical Safety**: Prioritize identifying critical or life-threatening findings. 3. **Methodical Workflow**: Follow standard radiological "Hanging Protocols" (Orientation -> Optimization -> Navigation -> Confirmation). --- ### AVAILABLE TOOLS > **[TOOL_DEFINITIONS_PLACEHOLDER]** > *(This section includes tools for loading data, metadata inspection, segmentation, visualization like MIP/Ortho/Montage, and radiomics analysis)*. --- ### CRITICAL THINKING RULES (DO NOT VIOLATE) * **Rule 0: NO-OBSERVATION, NO-VISUAL-CLAIMS**: Before the first successful imaging load, you MUST NOT claim any visual findings. * **Rule 1: THE VISUAL INTERPRETATION MANDATE**: Whenever you use a tool that generates a visualization, your IMMEDIATE next `thought` MUST explicitly interpret that visual artifact. * **REQUIRED**: "ANALYSIS: The histogram shows a secondary peak at 800+ HU. This confirms dense calcifications. PLAN: ...". * **Rule 2: THE "AVERAGE FALLACY" (STATISTICAL TRAP)**: When analyzing focal lesions within a larger organ, do NOT rely on the Mean HU of the entire organ mask. Check Max HU for calcifications or Min HU for fat/air. --- ### RADIOLOGICAL REASONING GUIDELINES (SOP) 1. **Phase 1: Orientation & Technical Quality**: Load study and verify scan range and modality. 2. **Phase 2: Image Optimization (Window/Level)**: Apply correct presets immediately, such as Lung (W:1500, L:-600), Soft Tissue(W:400, L:40), or Bone (W:2000, L:500). 3. **Phase 3: Anatomical Triangulation**: If a lesion is suspected in Axial, request Coronal/Sagittal views to confirm. --- ### REACT INSTRUCTIONS (STRICT EXECUTION ORDER) Every time you receive an Observation, follow this 2-step process: 1. **STEP 1: ANALYZE THE OBSERVATION**: * **Decode**: What did the tool return? * **Analyze**: Describe features or check for data outliers. * **Verdict**: Is the finding [NORMAL], [SUSPICIOUS], or [INCONCLUSIVE]? 2. **STEP 2: PLAN THE NEXT MOVE**: Decide the next tool (Zoom, Measure, Re-window, or Conclude). --- ### RESPONSE FORMAT You must respond STRICTLY in JSON format. The `thought` string MUST follow the structure: `"ANALYSIS: [Interpret previous data] ... PLAN: [Justify next action]"`. #### Scenario A: Tool Call (Action) ```json "type": "action", "thought": "ANALYSIS: ... PLAN: ...", "action": "name": "tool_name", "arguments": "param": "value" ``` #### Scenario B: Sufficient Information (Final Answer) ```json "type": "final_answer", "thought": "ANALYSIS: ... PLAN: ...", "answer": "Clinical explanation... \ Short Categorical Label or Value" ``` --- ### FINAL ANSWER RULES (STRICT ENFORCEMENT) 1. **THE BOXED CONSTRAINT**: Content inside ` ` must be a Short Categorical Label or a Single Value. No full sentences. 2. **SEMANTIC VS. NUMERIC OUTPUT**: * **Location**: Translate coordinates into Anatomical Terms (e.g., "posterior segment of the right upper lobe"). NEVER output raw x, y, z coordinates. * **Computational**: Output the specific calculated number with units (e.g., "14.5 c"). Figure 5: Structure of the System Prompt. The core principles, critical thinking rules, and standard operating procedures (SOPs) for the AI Medical Imaging Assistant are displayed. To facilitate a clear presentation, the specific technical definitions of available tools have been truncated. Table 4: Performance comparison on the 3D Rad 1.2k dataset. M3D natively supports 3D volumes, whereas other baseline models process individual 2D slices. Models denoted with w CT Lens leverage our proposed CTLens framework to integrate spatial-contextual information. Model Task 1Task 2Task 3Task 4 Task 5 Task 6 Avg. BLEU RougeBertBLEU RougeBertBLEU RougeBertaccaccacc Gemini-3-Pro11.7215.8173.3012.4719.3675.3512.0512.0453.3061.0042.5063.0055.50 w CT-Flow26.8538.2987.9425.9534.3887.6735.9834.1092.7860.5052.2675.0062.59 (↑15.13) (↑22.48) (↑14.64) (↑13.48) (↑15.02) (↑12.32) (↑23.93) (↑22.06) (↑39.48) (↓0.50)(↑9.76) (↑12.00) (↑7.09) Gpt-5.29.5116.6783.7710.7219.5284.341.085.0381.3759.5036.0070.0055.17 w CT-Flow19.6825.3884.6020.4226.6185.0826.1426.1986.9470.0043.0077.5063.50 (↑10.17) (↑8.71)(↑0.83)(↑9.70)(↑7.09)(↑0.74) (↑25.06) (↑21.16) (↑5.57) (↑10.50) (↑7.00)(↑7.50)(↑8.33) Claude-Opus5.1914.8582.435.7216.0383.626.6519.6783.0154.0014.5057.5042.00 w CT-Flow16.2525.3884.9216.7324.3984.4927.4530.6787.7260.8037.6965.9954.83 (↑11.06) (↑10.53) (↑2.49) (↑11.01) (↑8.36)(↑0.87) (↑20.80) (↑11.00) (↑4.71)(↑6.80) (↑23.19) (↑8.49) (↑12.83) Qwen3-VL-235B 21.1525.8885.5115.5819.2185.243.606.3984.4755.5026.0051.0044.17 w CT-Flow22.4324.6884.5621.4822.9884.1817.7520.1987.5858.3338.5065.7954.21 (↑1.28)(↓1.20)(↓0.95)(↑5.90)(↑3.77)(↓1.06) (↑14.15) (↑13.80) (↑3.11)(↑2.83) (↑12.50) (↑14.79) (↑10.04) GLM4.6-V18.2626.6486.2919.0025.4186.199.5414.8189.1363.6413.0763.1646.62 w CT-Flow23.2226.9881.2722.9026.1779.5723.5823.2683.9455.5037.8462.7752.04 (↑4.96)(↑0.34)(↓5.02)(↑3.90)(↑0.76)(↓6.62) (↑14.04) (↑8.45)(↓5.19)(↓8.14) (↑24.77) (↓0.39)(↑5.41) Qwen3-VL-8B16.4820.0980.5215.4818.7180.502.134.7082.0958.0025.5057.5047.00 w CT-Flow18.9723.9878.5317.9722.0172.1725.7420.9589.5039.1342.3965.6649.06 (↑2.49)(↑3.89)(↓1.99)(↑2.49)(↑3.30)(↓8.33) (↑23.61) (↑16.25) (↑7.41) (↓18.87) (↑16.89) (↑8.16)(↑2.06) fine-tuned26.8436.5788.2223.4234.5086.4820.8124.2859.8975.4054.0078.9969.46 (↑10.36) (↑16.48) (↑7.70)(↑7.94) (↑15.79) (↑5.98) (↑18.68) (↑19.58) (↓22.20) (↑17.40) (↑28.50) (↑21.49) (↑22.46) M3D12.1519.0484.9811.8020.6885.4613.0519.3990.5422.5027.0023.0024.17 3D-RAD26.3334.8389.4131.6639.8989.9031.2937.4694.5872.5036.0065.5058.00 Table 5: Performance comparison across different sce- narios. Models denoted with w CT-Flow show the inte- gration of spatial-contextual information. ModelScenario 1 Scenario 2 Scenario 3Ave. Gemini-3-Pro35.0035.0031.0033.67 w CT-Flow45.0043.0044.0044.00 (↑10.00)(↑8.00)(↑13.00)(↑10.33) Gpt-5.226.0036.0037.0033.00 w CT-Flow35.0040.0037.0037.33 (↑9.00)(↑4.00)(0.00)(↑4.33) Claude43.0039.0034.0038.67 w CT-Flow44.0044.0043.0043.67 (↑1.00)(↑5.00)(↑9.00)(↑5.00) Qwen3-VL-235B33.0030.0026.0029.67 w CT-Flow30.0036.0036.0034.00 (↓3.00)(↑6.00)(↑10.00)(↑4.33) GLM4.6-V25.0024.0027.0025.33 w CT-Flow25.0033.0034.0030.67 (0.00)(↑9.00)(↑7.00)(↑5.33) Qwen3-VL-8B31.0033.0022.0028.67 w CT-Flow30.0026.0020.0025.33 (↓1.00)(↓7.00)(↓2.00)(↓3.33) Fine-tuned43.0039.0046.0042.67 (↑12.00)(↑6.00)(↑24.00)(↑14.00) M3D17.0017.0017.0017.00 3D-RAD39.0034.0035.0036.00