Brain Tumor Classification + RAG
MRI brain tumor classifier (4 classes) combined with a retrieval-augmented Q&A system over a curated medical knowledge base.
Overview
Pairs image classification with retrieval so that a prediction comes with relevant, source-attributed background information instead of a bare probability.
My role
Applied ML Researcher
Problem
MRI classifiers output abstract probability scores (e.g., 98.4% Glioma) without medical explanation, making it difficult for clinical teams to verify the diagnostic rationale.
Approach
A modified ResNet18 classifies 224×224 MRI images into four classes. A knowledge base of 10+ medical documents is embedded with all-MiniLM-L6-v2 and indexed in FAISS (cosine similarity); retrieved passages feed a Hugging Face model (DialoGPT-medium) to generate answers. Flask endpoints cover prediction, prediction + RAG, and free-form Q&A, with Docker support.
Results
- ~95% classification accuracy on the test set.
- Every RAG answer is returned together with its source documents.