A Guideline-Aware AI Agent for Zero-Shot Target Volume Auto-Delineation

This paper introduces OncoAgent, a novel guideline-aware AI agent that achieves zero-shot, training-free auto-delineation of clinical target volumes by converting textual clinical guidelines into 3D contours, demonstrating superior adaptability and physician preference over traditional supervised deep learning models.

The Gist

Imagine you are a master chef trying to bake a perfect cake. In the world of radiation therapy, the "cake" is the tumor, and the "icing" is the radiation dose. The doctor's job is to draw a very precise line around the tumor (the Clinical Target Volume) to make sure they zap the cancer but leave the healthy organs (like the heart or lungs) untouched.

Traditionally, doctors had to draw this line by hand, which is slow and tiring. Later, computers learned to do it by studying thousands of examples drawn by experts. But here's the problem: Medical guidelines change. If the "recipe" for the cake changes (e.g., "now we need 2mm more icing on the left side"), the old computer program doesn't know what to do. You'd have to teach it all over again from scratch, which takes months and costs a fortune.

Enter OncoAgent: The "Smart Sous-Chef"

This paper introduces a new AI called OncoAgent. Instead of being a student who memorizes thousands of old drawings, OncoAgent is like a super-smart sous-chef who can read the recipe book and figure out the cake on the fly.

Here is how it works, broken down into simple concepts:

1. The "Read-It-and-Do-It" Magic (Zero-Shot Learning)

Most AI models are like parrots; they repeat what they've heard before. If you ask them to do something new, they get confused.

• OncoAgent is different. It is an "AI Agent." Think of it as a chef who has a rulebook (the clinical guidelines) written in plain English.
• When a new patient arrives, OncoAgent reads the rulebook: "Okay, the rule says: Take the tumor, add 10mm of space, but stop if you hit the lung."
• It doesn't need to have seen this specific patient before. It just reads the rules and builds the plan instantly. This is called "Zero-Shot" because it does the job without needing a "practice run" or retraining.

2. The "Lego Builder" Approach

OncoAgent doesn't just guess; it follows a strict, step-by-step logic, like building with Lego bricks:

• Step 1: It finds the tumor (GTV).
• Step 2: It uses a "magnifying glass" tool to expand the tumor by the exact amount the rulebook says (e.g., 10mm).
• Step 3: It uses a "cookie cutter" tool to carve away any part that touches a dangerous organ (like the heart).
• Step 4: It adds a final safety layer (PTV) to account for breathing.

Because it follows the text instructions literally, it never forgets a rule.

3. The "Instant Update" Superpower

This is the biggest game-changer.

• Old AI: If the medical guidelines change tomorrow, the old AI is useless until you spend months feeding it new data.
• OncoAgent: If the guidelines change, you just update the text file with the new rules. OncoAgent reads the new text and immediately starts following the new rules. No retraining, no waiting. It's like swapping a recipe card in a cookbook; the chef knows exactly what to do next.

The Results: Did it work?

The researchers tested OncoAgent on patients with esophageal cancer and compared it to the best existing AI (which had been trained on thousands of examples).

• Accuracy: OncoAgent was just as good as the super-trained AI. It got the shape of the tumor almost perfectly.
• Safety: In fact, doctors preferred OncoAgent! Why? Because the old AI sometimes got "sloppy" and drew the line too close to the heart or lungs. OncoAgent, being a strict rule-follower, never made those dangerous mistakes.
• Speed: It did the whole job in about 1.7 minutes.

The Bottom Line

Think of OncoAgent as a transparent, rule-following robot that replaces the "black box" AI.

• Old AI: A black box that guesses based on memory. If the rules change, it breaks.
• OncoAgent: A clear, logical assistant that reads the manual and builds the plan. If the manual changes, it adapts instantly.

This technology promises a future where radiation treatment planning is faster, safer, and can adapt to new medical discoveries in seconds rather than years. It turns the complex, rigid world of medical guidelines into a flexible, intelligent conversation between a doctor and a machine.

Technical Summary

Here is a detailed technical summary of the paper "A Guideline-Aware AI Agent for Zero-Shot Target Volume Auto-Delineation."

1. Problem Statement

In radiation therapy, defining the Clinical Target Volume (CTV) and Planning Target Volume (PTV) is a critical yet labor-intensive process. It requires manual delineation by radiation oncologists to encompass microscopic disease while sparing Organs at Risk (OARs).

• Current Limitations: Traditional Deep Learning (DL) models (e.g., nnU-Net) automate this but rely on massive datasets of expert-annotated ground truth. They function as "black boxes" and are inflexible: any update to clinical guidelines (e.g., margin changes, new protocols) requires costly re-annotation of data and model retraining.
• The Gap: There is a need for a system that can interpret natural language clinical guidelines directly and generate 3D contours without requiring task-specific training data, ensuring adaptability and interpretability.

2. Methodology: OncoAgent Framework

The authors propose OncoAgent, a novel AI agent framework that converts textual clinical guidelines into 3D target volumes in a zero-shot manner (no training on delineation data required).

Core Architecture

The system operates in two phases: Planning and Execution, driven by a Large Language Model (LLM) acting as a reasoning engine.

1. System Prompting Strategies:

   • Clinician Role Assignment: The LLM is assigned the persona of an expert radiation oncologist. It is provided with fixed computational templates for target volume definitions (e.g., $CTV = (GTV \oplus m_{ctv}) \setminus OARs$).
   • Guideline Parameterization: The prompt helps the LLM map variable terminology in guidelines to specific anatomical model names (e.g., "Lung" $\to$ Lung_L, Lung_R) and resolve margin ranges (e.g., "5–10 mm") based on patient context.
   • Robust Tool-Call Sequencing: The LLM generates a deterministic, ordered sequence of tool calls (JSON format) to avoid errors. It follows a strict logic: Query $\to$ Segment OARs $\to$ Union OARs $\to$ Dilate GTV $\to$ Subtract OARs $\to$ Dilate CTV $\to$ PTV.

2. Execution Engine:

   • Tool Invocation: The validated plan is executed by dispatching calls to:
     • Segmentation Tools: Pre-trained deep learning models to delineate anatomical structures (OARs) on patient CT scans.
     • Geometric Tools: Morphological operations (dilation, erosion) and Boolean operations (union, subtraction) using the VTK library.
   • Self-Refinement: If the generated plan violates the tool-call schema, the LLM iteratively self-corrects the plan before execution.
   • Post-Processing: Contour smoothing and hole-filling are applied to the final output.

3. Key Contributions

1. Training-Free Architecture: A general, two-phase agentic framework that converts free-text guidelines into 3D volumes using pre-trained tools, eliminating the need for expert-annotated CTV training data.
2. Performance Parity with Supervised Models: Demonstrated that a zero-shot agent can achieve performance comparable to state-of-the-art fully supervised models (nnU-Net) while ensuring strict adherence to anatomical safety rules.
3. Instant Adaptability & Generalization: Validated that the framework can instantly adapt to alternative clinical guidelines (e.g., CROSS Trial, JASTRO 2024) and generalize to new anatomical sites (e.g., prostate) without retraining, simply by updating the text input.

4. Experimental Results

The study was evaluated on 40 patients with mid-thoracic esophageal cancer.

• Dataset: 32 cases for training DL baselines; 8 cases for zero-shot testing.
• Baselines: nnU-Net, DDAU-Net, and nnU-Net (GTV Prior).
• Quantitative Metrics (CTV):
  • OncoAgent: Dice Similarity Coefficient (DSC) = 0.842, Mean Surface Distance (MSD) = 1.06 mm.
  • nnU-Net (GTV Prior): DSC = 0.862, MSD = 0.95 mm.
  • Result: OncoAgent showed no statistically significant difference ($p \ge 0.05$) compared to the best supervised baseline, despite having zero training data for the specific task.
• Clinical Assessment (Blinded Survey):
  • Two senior radiation oncologists rated OncoAgent significantly higher than the supervised baseline in Guideline Compliance (4.06 vs. 3.56), Modification Effort (4.12 vs. 3.44), and Clinical Acceptability.
  • Qualitative Analysis: OncoAgent strictly avoided OAR violations (e.g., lung invasion) due to explicit Boolean subtraction, whereas DL models tended to over-expand. While OncoAgent was slightly "over-eroded" near the heart, it was deemed safer.
• Zero-Shot Transferability:
  • Achieved high Tool Call F1 scores (0.73–1.00) when switching to different esophageal guidelines and a prostate guideline without retraining.

5. Significance and Impact

• Paradigm Shift: Moves medical AI from "learning from examples" to "reasoning from guidelines," offering a transparent and auditable workflow.
• Regulatory & Clinical Trust: The intermediate planning step (human-readable JSON plan) allows clinicians to review and approve the logic before contour generation, addressing the "black box" issue of DL.
• Scalability: Solves the critical bottleneck of guideline obsolescence. As clinical protocols evolve, OncoAgent can be updated instantly via text prompts, whereas DL models require expensive retraining cycles.
• Future Directions: The authors note limitations regarding reliance on OAR segmentation quality, potential LLM hallucinations, and the need for multi-center validation, but the framework establishes a scalable pathway for interpretable radiotherapy planning.