MAIRA-2 CXR Report Generator

This project implements a Gradio-based application that utilizes the MAIRA-2 model from Microsoft to generate radiology reports from chest X-ray images. Users can input URLs for frontal and lateral chest X-ray images, along with relevant clinical information, to automatically generate findings.

Features

• Image Input: Accepts URLs for frontal and lateral chest X-ray images.
• Clinical Context: Enables input of clinical indications, comparison studies, and technical details of the X-ray.
• Automated Report Generation: Leverages the MAIRA-2 model to generate findings based on the provided information.
• Gradio Interface: Provides an intuitive web interface for effortless interaction. Application Screenshot

Requirements

• Python 3.7 or higher
• The following Python packages (see requirements.txt):
  • transformers
  • torch
  • huggingface_hub
  • requests
  • Pillow (PIL)
  • gradio
  • sentencepiece
  • protobuf

Installation

1. Clone the repository:

   git clone https://github.com/mvdiogo/MAIRA-2-CXR-Report-Generator.git
   cd MAIRA-2-CXR-Report-Generator 
   

2. Create a virtual environment (optional but recommended):

   python -m venv venv
   source venv/bin/activate  # On Linux/macOS
   venv\Scripts\activate  # On Windows
   

3. Install the required packages:

   pip install -r requirements.txt
   pip install --force-reinstall git+https://github.com/huggingface/transformers.git@88d960937c81a32bfb63356a2e8ecf7999619681
   

4. Install PyTorch:

   Refer to the official PyTorch website for installation instructions specific to your system: https://pytorch.org/get-started/locally/

Usage

1. Hugging Face Token:

   To access the gated MAIRA-2 model, you'll need a Hugging Face token and permission.

   • Request Access: Go to https://huggingface.co/microsoft/maira-2/tree/main and request access to the model.

   • Generate Token: Create a Hugging Face token with read access at https://huggingface.co/settings/tokens.

   • Set Environment Variable: Set the HF_TOKEN environment variable (replace "your_huggingface_token" with your actual token):

     export HF_TOKEN="your_huggingface_token"  # Linux/macOS
     set HF_TOKEN="your_huggingface_token"       # Windows
     

     Important: This is for development/testing only. For production environments, use secure methods for storing and accessing credentials. Never commit your token to public repositories.

2. Run the application:

   python main.py
   

3. Open the URL provided in the console in your web browser. This will typically be something like http://127.0.0.1:7860 or a gradio.live URL.

4. Enter the URLs for the frontal and lateral chest X-ray images, along with the indication, comparison, and technique.

5. Click the "Generate Findings" button to generate the radiology report findings.

Notes

• The quality of the generated reports depends on the quality of the input images and the capabilities of the MAIRA-2 model. Experiment with different prompts and settings to optimize results.

Disclaimer

This project is for demonstration and educational purposes only and should not be used for clinical decision-making. Always consult with qualified healthcare professionals for medical advice.