🤖 MedRAX: Medical Reasoning Agent for Chest X-ray

Abstract

Chest X-rays (CXRs) play an integral role in driving critical decisions in disease management and patient care. While recent innovations have led to specialized models for various CXR interpretation tasks, these solutions often operate in isolation, limiting their practical utility in clinical practice. We present MedRAX, the first versatile AI agent that seamlessly integrates state-of-the-art CXR analysis tools and multimodal large language models into a unified framework. MedRAX dynamically leverages these models to address complex medical queries without requiring additional training. To rigorously evaluate its capabilities, we introduce ChestAgentBench, a comprehensive benchmark containing 2,500 complex medical queries across 7 diverse categories. Our experiments demonstrate that MedRAX achieves state-of-the-art performance compared to both open-source and proprietary models, representing a significant step toward the practical deployment of automated CXR interpretation systems.

MedRAX

MedRAX is built on a robust technical foundation:

• Core Architecture: Built on LangChain and LangGraph frameworks
• Language Model: Uses GPT-4o with vision capabilities as the backbone LLM
• Deployment: Supports both local and cloud-based deployments
• Interface: Production-ready interface built with Gradio
• Modular Design: Tool-agnostic architecture allowing easy integration of new capabilities

Integrated Tools

• Visual QA: Utilizes CheXagent and LLaVA-Med for complex visual understanding and medical reasoning
• Segmentation: Employs MedSAM and PSPNet model trained on ChestX-Det for precise anatomical structure identification
• Grounding: Uses Maira-2 for localizing specific findings in medical images
• Report Generation: Implements SwinV2 Transformer trained on CheXpert Plus for detailed medical reporting
• Disease Classification: Leverages DenseNet-121 from TorchXRayVision for detecting 18 pathology classes
• X-ray Generation: Utilizes RoentGen for synthetic CXR generation
• Utilities: Includes DICOM processing, visualization tools, and custom plotting capabilities
  
  

ChestAgentBench

We introduce ChestAgentBench, a comprehensive evaluation framework with 2,500 complex medical queries across 7 categories, built from 675 expert-curated clinical cases. The benchmark evaluates complex multi-step reasoning in CXR interpretation through:

• Detection
• Classification
• Localization
• Comparison
• Relationship
• Diagnosis
• Characterization

Download the benchmark: ChestAgentBench on Hugging Face

huggingface-cli download wanglab/chestagentbench --repo-type dataset --local-dir chestagentbench

Unzip the Eurorad figures to your local MedMAX directory.

unzip chestagentbench/figures.zip

To evaluate with GPT-4o, set your OpenAI API key and run the quickstart script.

export OPENAI_API_KEY="<your-openai-api-key>"
python quickstart.py \
    --model chatgpt-4o-latest \
    --temperature 0.2 \
    --max-cases 2 \
    --log-prefix chatgpt-4o-latest \
    --use-urls

Installation

Prerequisites

• Python 3.8+
• CUDA/GPU for best performance

Installation Steps

# Clone the repository
git clone https://github.com/bowang-lab/MedRAX.git
cd MedRAX

# Install package
pip install -e .

Getting Started

# Start the Gradio interface
python main.py

or if you run into permission issues

sudo -E env "PATH=$PATH" python main.py

You need to setup the model_dir inside main.py to the directory where you want to download or already have the weights of above tools from Hugging Face. Comment out the tools that you do not have access to. Make sure to setup your OpenAI API key in .env file!

Tool Selection and Initialization

MedRAX supports selective tool initialization, allowing you to use only the tools you need. Tools can be specified when initializing the agent (look at main.py):

selected_tools = [
    "ImageVisualizerTool",
    "ChestXRayClassifierTool",
    "ChestXRaySegmentationTool",
    # Add or remove tools as needed
]

agent, tools_dict = initialize_agent(
    "medrax/docs/system_prompts.txt",
    tools_to_use=selected_tools,
    model_dir="/model-weights"
)

Automatically Downloaded Models

The following tools will automatically download their model weights when initialized:

Classification Tool

ChestXRayClassifierTool(device=device)

Segmentation Tool

ChestXRaySegmentationTool(device=device)

Grounding Tool

XRayPhraseGroundingTool(
    cache_dir=model_dir, 
    temp_dir=temp_dir, 
    load_in_8bit=True, 
    device=device
)

• Maira-2 weights download to specified cache_dir
• 8-bit and 4-bit quantization available for reduced memory usage

LLaVA-Med Tool

LlavaMedTool(
    cache_dir=model_dir, 
    device=device, 
    load_in_8bit=True
)

• Automatic weight download to cache_dir
• 8-bit and 4-bit quantization available for reduced memory usage

Report Generation Tool

ChestXRayReportGeneratorTool(
    cache_dir=model_dir, 
    device=device
)

Visual QA Tool

XRayVQATool(
    cache_dir=model_dir, 
    device=device
)

• CheXagent weights download automatically

MedSAM Tool

Support for MedSAM segmentation will be added in a future update.

Utility Tools

No additional model weights required:

ImageVisualizerTool()
DicomProcessorTool(temp_dir=temp_dir)

Manual Setup Required

Image Generation Tool

ChestXRayGeneratorTool(
    model_path=f"{model_dir}/roentgen", 
    temp_dir=temp_dir, 
    device=device
)

• RoentGen weights require manual setup:
  1. Contact authors: https://github.com/StanfordMIMI/RoentGen
  2. Place weights in {model_dir}/roentgen
  3. Optional tool, can be excluded if not needed
     

Configuration Notes

Required Parameters

• model_dir or cache_dir: Base directory for model weights that Hugging Face uses
• temp_dir: Directory for temporary files
• device: "cuda" for GPU, "cpu" for CPU-only

Memory Management

• Consider selective tool initialization for resource constraints
• Use 8-bit quantization where available
• Some tools (LLaVA-Med, Grounding) are more resource-intensive
  

Local LLMs

If you are running a local LLM using frameworks like Ollama or LM Studio, you need to configure your environment variables accordingly. For example:

export OPENAI_BASE_URL="http://localhost:11434/v1"
export OPENAI_API_KEY="ollama"

Optional: OpenAI-compatible Providers

MedRAX supports OpenAI-compatible APIs, allowing regional or local LLM providers to serve as alternative backends.

For example, to use Qwen3-VL via Alibaba Cloud DashScope, set the following environment variables:

export OPENAI_BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
export OPENAI_API_KEY="<your-dashscope-api-key>"
export OPENAI_MODEL="qwen3-vl-235b-a22b-instruct"

Star History

Authors

• Adibvafa Fallahpour¹²³⁴ * (adibvafa.fallahpour@mail.utoronto.ca)
• Jun Ma²³ *
• Alif Munim³⁵ *
• Hongwei Lyu³
• Bo Wang¹²³⁶

¹ Department of Computer Science, University of Toronto, Toronto, Canada
² Vector Institute, Toronto, Canada
³ University Health Network, Toronto, Canada
⁴ Cohere, Toronto, Canada
⁵ Cohere Labs, Toronto, Canada
⁶ Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, Canada

* Equal contribution

Citation

If you find this work useful, please cite our paper:

@misc{fallahpour2025medraxmedicalreasoningagent,
      title={MedRAX: Medical Reasoning Agent for Chest X-ray}, 
      author={Adibvafa Fallahpour and Jun Ma and Alif Munim and Hongwei Lyu and Bo Wang},
      year={2025},
      eprint={2502.02673},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.02673}, 
}

---

Made with ❤️ at University of Toronto, Vector Institute, and University Health Network