Enhancing Medical Imaging Diagnostics Through Large Language Model-Based Knowledge Integration

Background: Medical imaging diagnosis faces challenges in accuracy and efficiency due to heavy reliance on clinician expertise. This study aims to develop a medical imaging diagnosis assistance system (LLM-RAG-MID) that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) technology, enhancing diagnostic reliability for complex lesions through dynamic knowledge retrieval. Methods: The LLM-RAG-MID system combines text preprocessing, vectorization, similarity computation, and knowledge fusion modules, with Transformer models generating diagnostic suggestions. RAG dynamically retrieves external medical literature and case data to augment LLM outputs. Three clinical cases—intraspinal ependymoma, low-grade glioma, and cirrhosis with multiple intrahepatic nodules—were evaluated. Diagnostic performance was compared between pure LLM and LLM-RAG-MID configurations, with outcomes assessed by clinical experts. Results: LLM-RAG-MID improved diagnostic accuracy, comprehensiveness, and logical consistency compared to pure LLM. In the ependymoma case, the system accurately identified the tumor type and systematically excluded five differential diagnoses, increasing diagnostic coverage by 40%. Across all cases, diagnostic outcomes aligned closely with expert clinical judgments. Conclusions: The integration of LLM and RAG technologies effectively enhances medical imaging diagnosis by dynamically integrating multi-source knowledge. This study demonstrates the potential of AI-assisted systems in reducing diagnostic subjectivity and supporting clinical decision-making.