How vinith-44/image-based-radiology-report-generation-using-medgemma works
RADAR drafts structured radiology reports from X-ray, CT, or MRI images. The documented end-to-end flow preprocesses an image, derives graph-guided disease estimates (GHFE) and RAMT confidence, retrieves and guards supporting clinical evidence (RadioShield), then generates a report with MedGemma. RAMT also has a separate semi-supervised training workflow. The README and excerpts establish the main stages; some wiring among unsampled modules is documented rather than directly visible in source.
Main entry points: medgemma_report_generator.py Report pipeline · radioshield_rag.py RadioShield RAG · medgemma_engine.py MedGemma inference · ghfe_module.py GHFE features
Components 12
Image analysis
- GHFE features
src/ghfe_module.py - RAMT inference
src/medgemma_report_generator.py
Evidence retrieval
- RadioShield RAG
src/radioshield_rag.py - Clinical evidence index
src/radioshield_rag.py - NLI faithfulness guard
src/radioshield_rag.py
Report generation
- Report pipeline
src/medgemma_report_generator.py - MedGemma inference
src/medgemma_engine.py - Structured report
src/medgemma_report_generator.py
Model training
- RAMT training
src/ramt_module.py
Other components
- Radiology user
- Medical image
- Hugging Face Hub
Connections 11
- Radiology user → Report pipeline: submits case
README.md - Medical image → Report pipeline: provides image
README.md - Report pipeline → GHFE features: extracts features
README.md - Report pipeline → RAMT inference: estimates confidence
README.md - RAMT training → GHFE features: reuses model
src/ramt_module.py - Report pipeline → RadioShield RAG: requests evidence
src/medgemma_report_generator.py - RadioShield RAG → Clinical evidence index: retrieves chunks
src/radioshield_rag.py - RadioShield RAG → NLI faithfulness guard: checks faithfulness
src/radioshield_rag.py - Report pipeline → MedGemma inference: generates report
- MedGemma inference → Hugging Face Hub: loads model
src/medgemma_engine.py - Report pipeline → Structured report: formats output
src/medgemma_report_generator.py