Source-linked AI summary
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
TL;DR
Static, single-pass 3D CT approaches do not reflect the iterative tool-mediated workflow of clinical interpretation. CT-Flow uses MCP to orchestrate imaging tools, while CT-FlowBench supplies trajectory-based training and evaluation. The authors report state-of-the-art performance across diagnostic benchmarks and transparent, traceable reasoning.
Problem
Most LVLM-based CT methods treat volumes as static inputs, while clinical interpretation involves iterative review and specialized tools.
Source: §1 · p. 1Method
CT-Flow uses MCP to orchestrate specialized imaging tools, and CT-FlowBench provides instruction-tuning data for tool-mediated reasoning trajectories.
Source: §1 · p. 2, §4.3 · p. 5Results
Takeaways & Limitations
The authors present CT-Flow as a promising path toward more interactive and reliable AI-assisted medicine.
Source: §6 · p. 8The current model relies primarily on supervised fine-tuning and has not integrated reinforcement learning to optimize decision trajectories or reduce redundant tool calls.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
The introduction argues that static, end-to-end CT analysis poorly matches clinicians’ iterative interpretation workflow. CT-Flow pairs MCP-based tool orchestration with a trajectory-focused benchmark, and the authors report gains over medical LVLM baselines.
1 Introduction
- Existing LVLM methods treat CT volumes as static inputs, and pre-encoded representations can obscure fine anatomical details and subtle radiographic cues.The introduction contrasts this with clinicians’ iterative use of slice navigation, measurements, and specialized tools.Source: §1 · p. 2
- CT-Flow uses MCP to turn volumetric analysis into sequential decision-making, decomposing diagnostic queries into tool calls and grounding reasoning in tool-verified evidence.Its tools support visualization, region-specific segmentation, and quantitative radiomics extraction.Source: §1 · p. 2
- CT-FlowBench evaluates executable reasoning chains and intermediate decisions and tool use, rather than only final-answer correctness.Experts and agents iteratively co-design and refine queries and ground-truth trajectories.Source: §1 · p. 2
- CT-Flow achieves substantial gains on established 3D CT VQA benchmarks and CT-FlowBench, outperforming specialized medical LVLMs; fine-tuned models also show generalization and robustness.The authors frame these results as evidence for agentic reasoning with tool use.Source: §1 · p. 2
- The authors describe CT-Flow as providing transparent, traceable, clinically aligned reasoning compared with end-to-end LVLM baselines.The motivation is that subtle 3D patterns and measurement-dependent criteria can influence triage and treatment decisions.Source: §1 · pp. 1–2
2 Related Work
Related work spans static 3D volume encoding, slice-sequence methods, and medical agents that invoke clinical tools. The paper positions CT-Flow’s iterative tool orchestration against these approaches.
2 Related Work
- Native 3D and serialization-based methods capture global patterns, but may retain limited fine-grained detail or lose spatial fidelity.The cited examples include 3D ViT and spatial pooling approaches, as well as slice-sequence modeling and representative-slice selection.Source: §2 · p. 3
- MCP provides an open, standardized interface for LLM access to external tools, with prior work applying it to tool access, memory, and context management.The paper presents this protocol as a basis for connecting reasoning with specialized tool invocation.Source: §2 · p. 3
- Medical agent frameworks integrate diagnostic tools or visual grounding, but many treat tool invocation as fragmented actions rather than iterative 3D analysis.The related-work discussion names ChatCAD, Med-Agents, and MedRAX among these efforts.Source: §2 · pp. 3–4
3 Methodology
CT-Flow represents CT interpretation as an orchestrated workflow: a standardized MCP action space supports repeated reasoning, tool use, and evidence-based refinement.
3 Methodology
- CT-Flow aims to shift LVLMs from static prediction to clinical workflow orchestration through diagnostic planning, on-demand imaging tools, and iterative verification.The system is presented as an orchestrator for CT tasks.Source: §3 · p. 4
3.1 Standardizing Clinical Interface via MCP
- The MCP action space groups imaging operations into data ingestion, global navigation, detailed observation, and advanced analysis.These tool suites abstract low-level image processing and provide a composable basis for iterative, goal-oriented probing.Source: §3.1 · p. 4
3.2 Iterative Probing over ReAct
- For a clinical query, the orchestrator interleaves reasoning states, MCP actions, and returned observations, then refines its understanding through further targeted probing when evidence is insufficient.The final answer is synthesized from evidence accumulated across the trajectory.Source: §3.2 · p. 4
4 Dataset Construction
CT-FlowBench is designed to represent 3D chest CT interpretation as executable, tool-mediated reasoning trajectories. Its construction emphasizes complex multi-step cases and checks that trajectories are grounded in raw-volume observations and annotated diagnoses.
4 Dataset Construction
- CT-FlowBench represents clinical volumetric interpretation as tool-mediated trajectories with intermediate observations retrievable from the raw volume, rather than final labels alone.This design is intended to reflect interactive radiologist workflows and enable step-level verification.Source: §4 · p. 4
4.1 Data Source and Curation
- The benchmark curates CT-RATE cases for anatomical diversity, diagnostic richness, and quantitative assessment, prioritizing complex multi-step reasoning and excluding redundant or non-actionable cases.The paper notes that trajectory synthesis over the full CT-RATE corpus is computationally expensive.Source: §4.1 · pp. 4–5
4.2 Task Scenario Definitions.
- The benchmark defines three functional scenario types spanning fundamental perception, holistic reasoning, and evaluation of discriminative precision and logical consistency.The supplied description names the scenario purposes but does not provide their individual definitions.Source: §4.2 · p. 5
4.3 Trajectory Synthesis
- Execution-in-the-loop synthesis builds trajectories of interleaved thoughts, actions, and observations, with candidate paths explored by a teacher model and filtered for procedural consistency.The consistency criterion requires physically grounded, retrievable observations and a logical chain ending in the ground-truth diagnosis.Source: §4.3 · p. 5
- CT-FlowBench contains 300 evaluation instances and approximately 2,000 QA training instances, including executable trajectories and 3D-RAD training data.Trajectories are retained only when tool execution reproduces intermediate observations from raw volumes and final answers match annotations.Source: §4.3 · p. 5
5 Experiments
CT-Flow is evaluated on 3D-RAD and CT-FlowBench against broad model baselines, with tests of tool-mediated reasoning and tool-category ablations. Results show strong 3D-RAD performance, while CT-FlowBench exposes the added difficulty of multi-stage tool use and smaller models’ limits.
5.1 Experimental Protocols
- CT-Flow evaluation covers CT-FlowBench and a refined 3D-RAD VQA subset, sampling 200 instances per task and randomizing option order.Table 1 reports average performance across 3D-RAD subtasks and CT-FlowBench scores for QA, AM, and DD.Source: §5.1 · p. 5, Table 1 · p. 6
5.2 Comparison with SOTA Methods.
- 69.46% accuracy on 3D-RAD makes CT-Flow-8B state of the art, a +22.46% improvement over the slice-based baseline.Figure 3 also reports gains of +8.33% for GPT-5.2 and +12.83% for Claude-Opus.Source: §5.2 · p. 6, Figure 3 · p. 7, §5.3 · p. 7
5.3 Analysis of Tool-mediated Reasoning.
- 44.33% is the highest average accuracy on CT-FlowBench, where multi-stage tool use is harder and Qwen3-8B regresses by 3.34%.Gemini-3-Pro and GPT-5.2 instead gain +10.33% and +14.00%, respectively, on the benchmark.Source: §5.2 · p. 7, §5.3 · p. 7
5.4 Tool-use Performance & Ablations
- Tool interactions increase with task complexity, while GPT-5.2 has near-zero tool-name errors and minimal argument hallucinations.Qwen3-VL-8B-Instruct has 0.782 tool-name errors, indicating difficulty maintaining consistency in long-chain tool manipulation.Source: §5.4 · pp. 7–8, Table 2 · p. 8
- Removing tool categories 2–4 lowers diagnostic accuracy and raises format errors, while ablations identify losses in quantitative analysis, subtle-lesion sensitivity, and slice navigation.The study concludes that all four tool categories are essential to the clinical reasoning pipeline.Source: Figure 4 · p. 8, §5.4 · p. 8
6 Conclusion
CT-Flow uses MCP and CT-FlowBench to shift 3D CT analysis toward protocol-driven active probing. Its evaluations report state-of-the-art diagnostic performance and transparent, traceable reasoning aligned with radiological workflows.
6 Conclusion
- CT-Flow integrates MCP with trajectory-based CT-FlowBench so an orchestrator can dynamically invoke specialized clinical tools.The framework is presented as a shift from passive visual ingestion to protocol-driven Active Probing.Source: §6 · p. 8
- Empirical results report state-of-the-art performance across diagnostic benchmarks and a transparent, traceable reasoning process aligned with real-world radiological workflows.The paper frames this as a promising path toward more interactive and reliable AI-assisted medicine.Source: §6 · p. 8
Limitations
The authors identify supervised fine-tuning without reinforcement learning and higher inference latency as limitations of CT-Flow.
Limitations
- CT-Flow relies on supervised fine-tuning without reinforcement learning, and its higher inference latency may challenge time-sensitive scenarios such as acute stroke triage.The authors propose future work on trajectory optimization and efficiency methods including parallel processing or trajectory pruning.Source: Limitations · p. 9
Ethical Considerations
CT-Flow is framed as clinical decision support with auditable reasoning, but its tool actions and diagnostic suggestions still require radiologist review. The paper also describes data-protection practices for benchmark construction and deployment.
Clinical Oversight and Agency
- CT-Flow is a clinical decision support system, not an autonomous diagnostic entity.The authors require certified radiologists to review automated tool sequences and final diagnostic suggestions because incorrect tool use or tool-output interpretation remains possible.Source: Ethical Considerations · p. 9
Transparency and Accountability
- MCP traceability lets clinicians inspect which slices were visualized and which radiomic features were extracted.The paper calls this an auditable trail of how a diagnostic conclusion was reached.Source: Ethical Considerations · p. 9
Data Privacy and Provenance
- CT-FlowBench uses de-identified, publicly available CT datasets, excludes PHI, and is designed for secure local clinical deployment.The authors state that data processing follows standard medical privacy regulations, including HIPAA.Source: Ethical Considerations · p. 9
A.1 Provided Tools in CT-Flow
CT-Flow organizes its medical imaging tools into four functional modules intended to mirror a clinical analysis workflow, from data validation through prognostic modeling.
A.1 Provided Tools in CT-Flow
- The four tool modules follow a workflow from raw-data validation and holistic screening to local quantification and AI-driven prognostic modeling.The paper says this modular organization is intended to support reproducibility and scalability in clinical studies.Source: §A.1 · p. 11
- The workflow moves from raw-data validation to holistic screening before local quantification.Source: §A.1 · p. 11
- AI-driven prognostic modeling is the final stage in the described tool workflow.Source: §A.1 · p. 11
A.2 Details of Dataset Construction
CT-Flow tools add spatial context to 2D slice-based analysis, with reported gains across report-generation and diagnostic tasks and scenario evaluations. The benchmark construction specifies report categorization, evidence-grounded labels, tool trajectories, and spatially referenced tasks.
A.2.1 Multi-Agent Annotation and Taxonomic Categorization
- Reports are screened for numerical presence, anatomical diversity, and clinical uncertainty, then categorized as Type A atomic, Type B logical, or Type C autonomous.Reports with explicit diagnostic uncertainty or follow-up recommendations are prioritized as Type C.Source: §A.2.1 · p. 11
A.2.2 High-Precision Ground Truth Synthesis
- Ground-truth findings are restricted to explicit measurements, anatomical descriptors, or HU values, and each fact is mapped to verification tools.For example, pulmonary-nodule findings are linked to segmentation and maximum-diameter measurement functions, so evaluation covers execution paths as well as answers.Source: §A.2.2 · p. 12
A.2.3 Clinical Scenario Domain Configuration
- Tasks span quantitative precision, surgical mapping, and diagnostic synthesis, including RECIST/HU analysis, vascular relationships, and systemic staging.Source: §A.2.3 · p. 12
A.2.4 The “Blind Verification” Instruction Strategy
- Blind Verification removes explicit pathology labels and uses spatial landmark references so agents must locate targets with visualization and segmentation tools.Each task is constrained to the predefined tool library and paired with an expert SOP trajectory for evaluating tool-call logic, efficiency, and safety.Source: §A.2.4 · p. 12, §A.2.5 · p. 13
A.3.1 Bridging the 2D-to-3D Gap (Table 1)
- Gemini-3-Pro’s report-generation BLEU score rises from 11.72 to 26.85 and its Rouge score from 15.81 to 38.29 with CT-Flow tools.The authors attribute the gains to spatial continuity across slices; the table caption identifies the 3D Rad 1.2k comparison.Source: §A.3.1 · p. 13, Table 4 · p. 16
- Claude-Opus’s Task 5 accuracy improves by 23.19% with CT-Flow tools.The authors frame this task as requiring identification of volumetric features.Source: §A.3.1 · p. 13, Table 4 · p. 16
A.3 Comparative Analysis: 2D Slice-based vs. CT-Flow Tool Processing
- Tool-enhanced models integrate 3D spatial context, and the authors report that they significantly outperform M3D, a native 3D model, on the 3D Rad 1.2k evaluation.The paper summarizes tool-based performance as rivaling or exceeding natively 3D-aware models and approaching fine-tuned systems.Source: §A.3.1 · p. 13, §A.3 · p. 13, §A.3.3 · p. 14
A.3.2 Scenario-Based Robustness (Table 2)
- Gemini-3-Pro gains an average of 10.33 points across scenarios and reaches 45.00 in Scenario 1 with CT-Flow tools.The authors also report that the tool integration helps address slice-level ambiguity and scenario variance; the table compares performance across scenarios.Source: §A.3.2 · p. 13, Table 5 · p. 17