README.md

metadata

license: mit
title: Radiology Report Generation System
sdk: docker
emoji: 🏢
colorFrom: blue
colorTo: green

Radiology AI Report Generation System v2.0

AI-powered radiology report generation using BLIP-2 vision + Phi-2 LLM fine-tuned on real radiology reports.

What it does

• Upload any radiology image (X-ray, CT, MRI, Ultrasound, PET, Mammography, etc.)
• BLIP-2 reads the actual image and describes what it sees
• Phi-2 (fine-tuned on 227,835 real radiology reports) generates a structured report
• Outputs: FINDINGS, IMPRESSION, RECOMMENDATIONS + UMLS/ICD-10/SNOMED CT codes

Requirements

• Windows 10/11
• Python 3.10
• NVIDIA GPU with 8GB+ VRAM (RTX 4060 or better)
• CUDA 12.8
• 20GB free disk space (for models)

---

Quick Start

Step 1 — Setup (run once)

Double-click: setup.bat

This creates venv, installs all dependencies, verifies GPU.

Step 2 — Fine-tune (run once, ~35 minutes)

Double-click: run_training.bat

Downloads unsloth/Radiology_mini (227,835 real reports) and fine-tunes Phi-2.

Step 3 — Start the system

Terminal 1:

run_backend.bat

Terminal 2:

run_frontend.bat

Step 4 — Use the system

Open browser: http://localhost:7860

---

Manual commands (if .bat files don't work)

cd D:\radiology_ai_system
venv\Scripts\activate

# Backend
set USE_4BIT=true
set ENABLE_VISION_CAPTION=true
uvicorn backend.main:app --reload --host 0.0.0.0 --port 8000

# Frontend (new terminal)
venv\Scripts\activate
python frontend/ui.py

# Fine-tuning (Git Bash)
python training/train_unsloth.py --max_steps 300

---

Project Structure

radiology_ai_system/
├── backend/
│   ├── main.py              # FastAPI endpoints
│   ├── model_loader.py      # BLIP-2 + Phi-2 loading
│   ├── preprocessing.py     # Image preprocessing (9 modalities)
│   ├── report_generator.py  # Report generation + parsing
│   ├── validation_module.py # Report quality scoring
│   └── medical_coding.py    # UMLS/ICD-10/SNOMED CT
├── frontend/
│   └── ui.py                # Gradio UI
├── training/
│   ├── train_unsloth.py     # Fine-tuning on Radiology_mini
│   ├── prepare_dataset.py   # Dataset preparation
│   └── evaluate_model.py    # Model evaluation
├── models/
│   ├── slm_model/           # Phi-2 cached here
│   │   └── lora_adapter/    # Fine-tuned LoRA adapter saved here
│   └── vision_encoder/      # BLIP-2 cached here
├── setup.bat                # One-time setup
├── run_backend.bat          # Start backend
├── run_frontend.bat         # Start frontend
├── run_training.bat         # Fine-tune model
└── requirements.txt

---

VRAM Usage

Component	VRAM
Phi-2 (4-bit)	~2.0GB
BLIP-2	~3.5GB
Total	~5.5GB
Available	8.6GB ✅

---

API Endpoints

• POST /api/generate-report — Generate report from image
• GET /api/report/{id} — Get stored report
• GET /api/export/pdf/{id} — Download PDF
• GET /api/export/json/{id} — Download JSON
• GET /api/health — System health
• GET /docs — Swagger UI (http://localhost:8000/docs)