MedGemma POC

A proof-of-concept medical image analysis application powered by MedGemma 1.5 4B-it.

⚠️ EXPERIMENTAL RESEARCH TOOL ONLY

• Outputs may contain hallucinations, factual errors, or dangerous suggestions
• NOT VALIDATED FOR CLINICAL USE — Do not use for diagnosis, treatment, or patient care
• Follow HAI-DEF terms; attribute properly; validate rigorously before any real-world extension

Stack

Layer	Technology
Frontend	Bun + Vite + Svelte
Backend	UV + FastAPI + SQLite
ML	MedGemma 1.5 4B-it (via transformers)
Task Runner	just

Prerequisites

• Bun ≥1.0 — https://bun.sh
• UV ≥0.4 — https://docs.astral.sh/uv/
• Python 3.11+ (managed via UV)
• Hugging Face account with MedGemma access granted

Hardware Requirements

Setup	Memory	Notes
Apple Silicon (M1/M2/M3/M4)	16GB+ unified	Uses MPS backend
NVIDIA GPU	16GB+ VRAM	Or 8GB with 4-bit quantization
CPU only	32GB+ RAM	Very slow (~60s/image), not recommended

Quick Start

# 1. Clone and enter project
cd medgemma-poc

# 2. Run setup script
./scripts/setup.sh

# 3. Set your Hugging Face token
cp .env.example .env
# Edit .env and add your HF_TOKEN

# 4. Start development servers
just dev

Development Commands

just dev        # Start both frontend and backend
just frontend   # Start frontend only
just backend    # Start backend only
just test       # Run all tests
just lint       # Lint all code
just format     # Format all code
just clean      # Clean build artifacts

Project Structure

medgemma-poc/
├── frontend/         # Bun + Vite + Svelte
├── backend/          # UV + FastAPI
├── ml/               # ML inference, evaluation, preprocessing
├── data/             # Local data (gitignored)
├── .github/          # CI/CD workflows
├── docker/           # Container definitions
├── scripts/          # Setup and utility scripts
└── docs/             # Documentation

API Endpoints

Method	Endpoint	Description
POST	/api/analyze	Analyze medical image(s) with prompt
GET	/api/history	Get conversation history
DELETE	/api/history/{id}	Delete conversation
GET	/api/health	Health check

License

MIT — See LICENSE

Acknowledgments

• MedGemma by Google Health AI
• Built following HAI-DEF (Health AI Developer Foundations) guidelines