MIDAS - System

Multi-Modal Intelligent Dermoscopy Analysis System

An AI-powered skin cancer analysis platform combining a multi-agent RAG chat assistant for dermatology Q&A with CNN-based lesion image analysis.

Overview

MIDAS gives patients and clinicians two ways to get dermatology insight:

• Chat assistant — ask dermatology questions and get answers grounded in an indexed knowledge base, with web search fallback for out-of-index queries.
• Image analysis — upload a lesion photo and get a CNN classification, risk assessment, and a generated PDF report.

Access is role-based (patient, clinician, admin), enforced server-side on every route.

Tech Stack

• Frontend: Next.js 15 (App Router), Clerk for auth, TanStack Query, Zustand
• Backend: FastAPI (Python 3.11)
• Data: Supabase — Postgres + pgvector, Storage, Deno edge functions
• AI: NVIDIA-hosted LLM for chat, Tavily for web search, a PyTorch CNN for lesion classification
• Embeddings: all-MiniLM-L6-v2 (384-dim), with a B+ tree chunk index for retrieval

Redis is present in config and Docker but not currently wired up — nothing connects to it yet.

Architecture

Browser (Next.js + Clerk)
   │
   ▼
Next.js route handlers  (BFF — same-origin API, server-only backend URL)
   │
   ▼
FastAPI backend  (backend/app/main_agent.py)
   │
   ├─▶ Supabase (Postgres + pgvector + Storage)
   ├─▶ NVIDIA-hosted LLM
   └─▶ Tavily web search

Chat, image analysis, and report generation each follow this same browser → Next.js → FastAPI → Supabase shape.

Quick Start

Prerequisites

• Node.js 18+
• Python 3.11
• A Supabase project (Postgres + Storage configured)

Start everything

./start-services.sh     # backend on :8000, frontend on :3000
./verify-services.sh    # health checks
./stop-services.sh

This creates the backend virtualenv, installs dependencies, and sources the .env files automatically.

Or run manually

# Backend — must run from backend/ (config loads .env relative to CWD)
cd backend
source venv/bin/activate
python -m uvicorn app.main_agent:app --host 0.0.0.0 --port 8000 --reload

# Frontend (npm, not pnpm/yarn)
cd frontend
npm run dev

• Chat UI: http://localhost:3000/dashboard/assistant
• API docs: http://localhost:8000/docs
• Health check: http://localhost:8000/healthz

Project Structure

MIDAS_final_work/
├── frontend/          # Next.js 15 app (App Router, Clerk auth)
│   └── src/
├── backend/           # FastAPI app
│   ├── app/           # live entry point: app/main_agent.py
│   └── models/        # trained CNN weights
├── supabase/          # migrations + edge functions
├── docs/              # architecture and reference docs
├── tests/             # standalone test scripts
└── scripts/           # docker-compose wrapper scripts

Environment

The backend reads .env / .env.backend (repo root and backend/); the frontend reads the root .env plus frontend/.env.local. Required variables include Supabase credentials, Clerk keys, and API keys for the LLM and web search providers — see the example env files for the full list.

Never commit real secrets. The tracked .env*.example files show the expected shape only.

Testing

There's no pytest suite or frontend test runner — tests are standalone scripts that mostly require the backend running locally plus live Supabase credentials:

• tests/ — integration scripts (test_bplus_indexing.py, test_react_cot.py, test_chat_persistence.py, …)
• backend/test_*.py — run from backend/ with the venv active

The CI workflow (.github/workflows/ci.yml) is currently a placeholder and does not build, lint, or test anything — treat a green check as informational only. Local verification (npx tsc --noEmit, npm run lint, npm run build, manual test scripts) is what actually validates changes.

Docker

docker-compose up   # redis, backend (:8000), frontend (:3001), nginx (:8080)

scripts/{up,down,logs,rebuild}.sh are thin wrappers around the same compose file.

Deployment

Railway configs are included for both services (backend/railway.toml, frontend/railway.toml), and a root render.yaml supports deploying to Render.

License

MIT — see LICENSE.

Acknowledgments

• HAM10000 and PAD-UFES-20 datasets, used to train the lesion classification model