Dr. A.Shilpa Gupta^1,, M. Uday Reddy¹, K. Jayanth Kumar Reddy¹, P. Medha Goud¹, T. Harshitha Reddy¹, Aditi Jopat¹

We introduce MMedRAG++,a medical multimodal retrieval system that uses fusion-based representation learning and graph-based reranking (Graph-RAG) to improve on conventional Retrieval-Augmented Generation (RAG). Unlike baseline systems, which do not incorporate fusion strategies or graph-based reranking, MMedRAG++ improves cross-modal embeddings and retrieval coherence. Experiments are conducted primarily on PMC-OA, a challenging dataset with only ~10% unique captions and many unrelated image-text pairs, and IU-Xray is used for modality- specific subtasks. Graph-RAG demonstrates improved retrieval centrality and diversity, and fusion strategies, including Cross- Attention and DeepSet Fusion, enhance embedding quality. Quantitative evaluation on PMC-OA and IU-Xray confirms improved retrieval coherence and cross-modal alignment over baseline configurations. Top-1 Accuracy: 18.7%, Top-10 Accuracy: 57.4%.

Medical AI, RAG, Graph-RAG, Multimodal Fusion, Contrastive Learning, Medical Image-text Retrieval