🩺 AI Wound Classification & Medical RAG System

An end-to-end Flask-based AI application that classifies wound images using a EfficientNet (TensorFlow) and provides grounded medical guidance using RAG (Retrieval-Augmented Generation).

---

Dataset Link - https://drive.google.com/drive/folders/1vgbEtBshfLRfwxig2wKiTb3mss5_PQt2?usp=sharing Model Link - https://drive.google.com/file/d/1bzTvevRJpmOdZ8VDCxPOG6pu_Flsdo4n/view?usp=sharing

🔹 System Architecture

The application utilizes a modular AI pipeline:

1. Vision Pipeline: Processes image uploads through a TensorFlow CNN to identify 8 specific wound classes.
2. RAG Pipeline: Uses LangChain and ChromaDB to retrieve context from trusted medical PDFs.
3. Inference Engine: Generates safe, non-hallucinated responses via a local Ollama LLM.

---

🔹 Technology Stack

Layer	Technology	Role
Backend	Flask	API Orchestration & Routing
ML Model	TensorFlow	CNN-based Image Classification
RAG Framework	LangChain	Knowledge Retrieval Logic
Vector DB	Chroma	Semantic Search & Embeddings
Containerization	Docker	Environment Isolation & Portability
LLM	Ollama	Local Inference (LlamaMedicine)

---

🔹 Supported Wound Classes

• Abrasions
• Bruises
• Burns
• Cut
• Ingrown_nails
• Laceration
• Stab_wound
• Healthy

---

🔹 Project Structure

.
├── app.py                      # Main Flask Backend
├── Dockerfile                  # Container Configuration
├── wound_classifier_final.keras # Trained CNN Model
├── class_names.json            # Wound Label Mapping
├── medical_knowledge_db/       # Trusted Medical PDFs
├── chroma_db/                  # Persistent Vector Store
├── uploads/                    # User Uploaded Images
├── templates/                  # Frontend UI (Jinja2)
└── requirements.txt            # Dependencies

🔹 Deployment (Docker)

This project is containerized for production-ready consistency. It uses a Hybrid Architecture where the application logic is isolated in Docker while connecting to the host machine's LLM service.
To Build docker build -t <image_name> . To Run docker run -d --name woundapp
-p 5001:5000
-e OLLAMA_BASE_URL="http://host.docker.internal:11434"
-v "$(pwd)/medical_knowledge_db:/app/medical_knowledge_db"
-v "$(pwd)/chroma_db:/app/chroma_db"
-v "$(pwd)/uploads:/app/uploads"
wound-rag-app

🔹 Technical Implementation Details:

• Networking: host.docker.internal allows the containerized Flask app to communicate with the Ollama service running on the host OS.
• Volumes: Persistent storage is mounted for the medical_knowledge_db and chroma_db to ensure search indices remain intact during restarts.
• Environment Variables: The OLLAMA_BASE_URL allows for flexible LLM endpoint configuration without modifying code.

🔹 Key Technical Highlights for Interviews

• RAG over Plain LLM: Prevents medical hallucinations by forcing the model to answer based only on provided medical literature.
• Edge Privacy: By using Ollama, the system performs local inference, ensuring sensitive medical data never leaves the local environment.
• Model Optimization: The CNN handles spatial feature extraction (texture/edges), while the RAG pipeline handles semantic knowledge retrieval.
• DevOps Readiness: Full Dockerization ensures the "it works on my machine" problem is eliminated, providing a clean path to cloud deployment.

🔹 Authors

1. Yuvaraj Kate (yuvarajkate1740@gmail.com)
2. Dhamkirti Sisodia (dksisodia002@gmail.com)