A Hybrid Retrieval-And-Generation Framework For Radiology Report Summarization With Faiss Indexing and T5 Transformers

Radiology reports contain clinically critical yet complex information that can overwhelm physicians when available in a raw narrative form. To address this challenge, this study proposes a Retrieval-Augmented Generation (RAG) approach to medical report summarization by integrating a FAISS-based semantic search engine with a small T5 generative model. The methodology uses Sentence-BERT to embed the “Findings” section of radiology reports from the MIMIC-III dataset, which are then indexed using FAISS to retrieve semantically similar cases. In this approach, The retrieved context has been merged with the original 'Finding' section and subsequently fed into the T5 model to produce the respective 'Impression' summary. Experimental results demonstrate the model’s effectiveness in terms of both lexical fidelity and semantic consistency, achieving ROUGE-1 of 0.5299, ROUGE-2 of 0.4206, METEOR of 0.5018, and compression ratio of 0.9213. Domain-specific semantic matching was validated using FactCC (0.5463), CheXpert Label Agreement (0.6194), and Medical Concept Overlap (0.5552) scores. The model demonstrated stable convergence over seven epochs without overfitting, as evidenced by the steadily decreasing validation loss. Qualitative examples show that the model generates smooth and clinically consistent summaries, although occasional factual hallucinations indicate areas for improvement. Overall, the proposed FAISS+T5 process addresses key limitations of traditional summarization methods by integrating contextual retrieval and generation. This framework offers a scalable, interpretable, and domain-specific solution for summarizing clinical texts and shows promise for implementation in real-world decision support systems.