MedGemma Impact Challenge Submission

🏥 Project: AI-Powered Radiology Assistant

An AI-powered radiology assistant that uses MedGemma to help radiologists analyze medical images, generate preliminary reports, and provide clinical decision support.

📋 Competition Overview

• Competition: The MedGemma Impact Challenge
• Sponsor: Google Research
• Prize Pool: $100,000
• Deadline: February 24, 2026

Prize Tracks

• Main Track: $75,000 (1st: $30k, 2nd: $20k, 3rd: $15k, 4th: $10k)
• Agentic Workflow Prize: $10,000 (Two $5k prizes)
• Edge AI Prize: $5,000
• Novel Task Prize: $10,000 (Two $5k prizes)

🎯 Submission Requirements

Mandatory Components

1. ✅ High-quality writeup (3 pages or less) - see docs/writeup.md
2. ✅ Reproducible code - this repository
3. ✅ Video demo (3 minutes or less) - to be created

Evaluation Criteria

🚀 Quick Start

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (recommended: 16GB+ VRAM for MedGemma 4B)
• Hugging Face account with HAI-DEF terms accepted

Installation

# Clone the repository
cd /home/david/Documents/medgemma

# Create virtual environment
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Accept HAI-DEF terms on Hugging Face
# Visit: https://huggingface.co/google/medgemma-4b-it

# Login to Hugging Face
huggingface-cli login

Running the Application

# Run the Gradio demo
python app.py

# Or run with Streamlit
streamlit run streamlit_app.py

📁 Project Structure

medgemma/
├── README.md                 # This file
├── requirements.txt          # Python dependencies
├── app.py                    # Main Gradio application
├── streamlit_app.py          # Alternative Streamlit app
├── src/
│   ├── __init__.py
│   ├── model.py             # MedGemma model wrapper
│   ├── agents.py            # Agentic workflow components
│   ├── utils.py             # Utility functions
│   └── prompts.py           # Prompt templates
├── notebooks/
│   ├── 01_quickstart.ipynb  # Quick start notebook
│   ├── 02_fine_tuning.ipynb # Fine-tuning example
│   └── 03_evaluation.ipynb  # Model evaluation
├── docs/
│   ├── writeup.md           # Competition writeup
│   └── technical_details.md # Technical documentation
├── tests/
│   └── test_model.py        # Unit tests
└── examples/
    └── sample_images/       # Sample medical images for demo

🔧 HAI-DEF Models Used

Primary Model

• MedGemma 4B IT (google/medgemma-4b-it): Multimodal model for medical image and text comprehension

Supporting Models (Optional)

• MedSigLIP (google/medsiglip-448): For medical image embeddings and retrieval
• MedASR (google/medasr): For voice-based radiology dictation
• MedGemma 27B (google/medgemma-27b-text-it): For complex clinical reasoning

📊 Use Cases Demonstrated

1. Chest X-Ray Analysis: Automated analysis of chest radiographs
2. Report Generation: AI-assisted radiology report drafting
3. Clinical Q&A: Interactive question-answering about medical images
4. Differential Diagnosis: AI-powered diagnostic suggestions

🔗 Resources

Official HAI-DEF Resources

• HAI-DEF Main Site
• MedGemma Collection
• HAI-DEF Collection
• Concept Apps
• Developer Forum

Notebooks & Examples

• Quick Start Notebook
• Fine-tuning Notebook
• MedGemma GitHub

⚠️ Important Disclaimers

• MedGemma is NOT intended for direct clinical use without proper validation
• All outputs require independent verification by qualified healthcare professionals
• This is a demonstration application for the hackathon competition
• Use of HAI-DEF models is subject to HAI-DEF Terms of Use

📝 License

This project is licensed under CC BY 4.0 as required by the competition rules.

📧 Contact

[Your team information here]

---

Built with MedGemma for the MedGemma Impact Challenge 2026