MedInsight: Evaluating Retrieval-Augmented Vision-Language Models for Evidence-Grounded Medical Image Understanding

MedInsight evaluates whether retrieval augmentation improves vision-language model performance for medical visual question answering. The study compares a baseline BLIP-2 model with a retrieval-augmented pipeline using CLIP, FAISS, and the ROCOv2 dataset, evaluated on the VQA-RAD benchmark with statistical significance testing. The project includes reproducible code, datasets, and experimental configurations.