CT-Flow: Orchestrating CT Interpretation Workflow with Model Context Protocol Servers

This paper introduces CT-Flow, an agentic framework that leverages the Model Context Protocol to transform static 3D CT analysis into a dynamic, tool-mediated workflow, achieving state-of-the-art performance on the newly curated CT-FlowBench by autonomously orchestrating complex diagnostic tasks through iterative tool use.

The Gist

Imagine you are trying to solve a complex 3D puzzle, like finding a tiny crack inside a massive, intricate crystal ball.

The Old Way (Traditional AI):
Most current medical AI models are like a student who is handed a stack of 200 flat photos of that crystal ball and told, "Find the crack." They look at the photos one by one, try to guess the answer, and spit out a result. They never actually touch the crystal ball. They can't zoom in, they can't measure the crack, and they can't rotate the ball to see the other side. If the crack is hidden in a specific angle, the AI might miss it entirely because it's just "guessing" based on a flat picture.

The New Way (CT-Flow):
The paper introduces CT-Flow, which is like giving that student a robotic assistant with a full toolkit. Instead of just looking at photos, this AI is now an active detective.

Here is how it works, broken down with simple analogies:

1. The "Model Context Protocol" (MCP) = The Universal Remote

Think of the AI as a smart TV. In the past, the TV could only show you a pre-recorded show (static images).
CT-Flow connects the TV to a Universal Remote (the Model Context Protocol). Now, the AI can press buttons to:

• Zoom in on a specific spot.
• Rotate the view to see the back.
• Measure the size of a tumor with a digital ruler.
• Run a chemical analysis (radiomics) to see what the tissue is made of.

2. The Workflow = A Detective's Investigation

When a doctor asks, "Is there a problem in this patient's lung?", the AI doesn't just guess. It follows a strict, step-by-step investigation plan, just like a human radiologist:

• Step 1: Orientation. "Okay, I need to load the patient's scan first." (The AI loads the 3D data).
• Step 2: Navigation. "I see a shadow here. Let me switch to a 'Lung Window' to see it better." (The AI changes the image settings).
• Step 3: Probing. "That shadow looks suspicious. Let me measure its exact size." (The AI uses a tool to measure).
• Step 4: Verification. "Is it fluid or solid? Let me check the density." (The AI runs a density analysis tool).
• Step 5: Conclusion. "Based on the measurements and the shape, this is likely a cyst, not a tumor."

3. The "CT-FlowBench" = The Final Exam

To make sure this new AI is actually good, the researchers built a special test called CT-FlowBench.

• Old Tests: Asked, "What is the answer?" (Multiple choice).
• CT-FlowBench: Asks, "Show me your work." It checks if the AI used the right tools, measured the right things, and followed a logical path to get the answer. It's like grading a math student not just on the final number, but on whether they showed their steps correctly.

Why Does This Matter?

• Accuracy: The paper shows that this "active detective" approach is 41% more accurate than the old "passive guesser" models.
• Trust: Because the AI has to "show its work" by using real tools (like measuring a tumor), doctors can trust the result more. They can see exactly how the AI reached its conclusion.
• Realism: It mimics how real doctors work. Doctors don't just stare at a screen; they scroll, zoom, measure, and compare. CT-Flow finally gives AI the ability to do the same.

The Bottom Line

CT-Flow changes medical AI from a passive observer (who just looks at pictures) into an active surgeon (who can pick up tools, measure, and investigate). It bridges the gap between a computer's raw processing power and the complex, hands-on reality of a doctor's office.

Technical Summary

1. Problem Statement

Current Large Vision-Language Models (LVLMs) for 3D Computed Tomography (CT) analysis suffer from a fundamental misalignment with real-world clinical workflows.

• Static vs. Dynamic: Existing approaches treat CT volumes as static visual inputs, relying on single-pass, end-to-end inference (e.g., using 3D Vision Transformers or serialized 2D slices). This creates information bottlenecks, obscuring fine-grained anatomical details and subtle radiographic cues.
• Lack of Agency: Clinical interpretation is inherently iterative. Radiologists actively scroll through slices, switch planes, measure lesions, and use segmentation or radiomics tools to refine hypotheses. Current LVLMs operate in a "read-only" mode, lacking the agency to perform these iterative, tool-mediated actions.
• Data Limitations: Existing benchmarks focus on final-answer correctness (Visual Question Answering) rather than the validity of the reasoning trajectory or tool usage, failing to evaluate the agent's ability to navigate complex 3D spaces.

2. Methodology: The CT-Flow Framework

The authors propose CT-Flow, an agentic framework that transforms passive volumetric encoding into active, tool-mediated probing using the Model Context Protocol (MCP).

A. Architecture & Tool Orchestration

CT-Flow decouples the LLM orchestrator from the imaging environment via MCP servers, creating a standardized interface for clinical utilities. The framework defines an atomic action space comprising four tool suites:

1. Data Ingestion: Loads CT volumes and metadata into a standardized, queryable 3D state.
2. Global Navigation: Enables fast whole-volume orientation and coarse anatomical localization (e.g., MIP, MinIP views).
3. Detailed Observation: Retrieves targeted high-resolution views (slices, orthogonal MPR, montages) to verify hypotheses.
4. Advanced Analysis: Performs quantitative measurements (Hounsfield Units), segmentation, and radiomics extraction.

B. Reasoning Paradigm (ReAct)

The diagnostic process is formalized as a Reasoning-Acting Trajectory ($T$). Instead of a single prediction, the model generates a sequential path of interleaved thoughts, actions, and observations:
$$T = \{(s_0, a_0, o_0), (s_1, a_1, o_1), \dots, (s_n, a_n, o_n)\}$$

• $s_t$ (Thought): The model's internal reasoning state for planning the next step.
• $a_t$ (Action): A specific tool call issued via MCP (e.g., view_ortho, segment_total_anatomy).
• $o_t$ (Observation): The visual or quantitative evidence returned by the tool.
  This iterative loop allows the model to refine its hypothesis based on tool-verified evidence, mimicking a radiologist's workflow.

C. Dataset Construction: CT-FlowBench

To support this paradigm, the authors introduce CT-FlowBench, the first benchmark tailored for 3D CT tool-use and multi-step reasoning.

• Source: Built upon the CT-RATE corpus, curated for high reasoning density (anatomical diversity, quantitative potential).
• Synthesis Strategy: Uses an "Execution-in-the-loop" approach with teacher models (e.g., GPT-4o, Gemini) to generate reasoning trajectories.
• Validation: Trajectories are filtered via a Procedural Consistency criterion, ensuring every tool action produces a physically retrievable observation from the raw volume and that the logical chain leads to the ground-truth diagnosis.
• Scenarios: Includes Quantitative Analysis, Spatial Mapping, and Diagnostic Inference.
• Scale: ~2,000 training instances and 300 evaluation instances.

3. Key Contributions

1. CT-Flow Framework: The first agentic architecture leveraging MCP to shift 3D medical analysis from passive encoding to active, tool-mediated probing, aligning AI behavior with clinical workflows.
2. CT-FlowBench: A novel benchmark that evaluates agent trajectories (intermediate decisions and tool usage) rather than just final answers, providing a standardized testbed for radiological agentic reasoning.
3. Performance & Transparency: Demonstrates that reframing CT understanding as agentic reasoning yields state-of-the-art (SOTA) performance while producing transparent, traceable, and clinically aligned reasoning processes.

4. Experimental Results

Experiments were conducted on CT-FlowBench and the 3D-RAD dataset.

• Diagnostic Accuracy:
  • CT-Flow-8B (fine-tuned) achieved 69.46% accuracy on 3D-RAD, surpassing the base model by +22.46% and outperforming specialized medical models like M3D-RAD (58.00%).
  • On CT-FlowBench, CT-Flow-7B/8B achieved average accuracies of 44.33% and 43.00% respectively, significantly outperforming non-agentic baselines.
• Generalization: General-purpose frontier models (e.g., GPT-5.2, Gemini-3-Pro) equipped with CT-Flow tools significantly outperformed specialized medical models, proving that tool-mediated reasoning can bridge the domain expertise gap without exhaustive pre-training.
• Tool-Use Reliability:
  • CT-Flow models demonstrated high reliability with near-zero tool-name errors and minimal argument hallucinations.
  • Smaller models (e.g., Qwen3-8B) showed performance regression on complex benchmarks, highlighting the need for sufficient reasoning capacity to manage autonomous tool-calling chains.
• Ablation Studies: Removing any of the four tool categories (Data Ingestion, Navigation, Observation, Analysis) resulted in significant drops in accuracy, confirming the necessity of the full hierarchical toolset.

5. Significance and Impact

• Paradigm Shift: Moves the field from "static prediction" to "dynamic orchestration," acknowledging that medical diagnosis is a process of active inquiry rather than passive observation.
• Clinical Alignment: By mimicking the radiologist's workflow (navigation -> measurement -> verification), the system produces traceable reasoning paths. Clinicians can audit which slices were viewed and what measurements were taken, addressing the "black box" issue of traditional deep learning.
• Scalability: The use of MCP allows for the easy integration of new clinical tools without retraining the core LLM, offering a scalable path for integrating autonomous intelligence into real-world radiology.
• Future Directions: The authors note limitations in inference latency and the potential for future integration of Reinforcement Learning (RL) to optimize decision trajectories and reduce redundant tool invocations.

In conclusion, CT-Flow establishes a new standard for 3D medical AI, demonstrating that agentic frameworks with standardized tool access are superior to static end-to-end models for complex, evidence-based diagnostic tasks.