XMedFusion: A Knowledge-Guided Multimodal Perception and Reasoning Framework for Autonomous Medical Systems

XMedFusion is a new modular AI framework designed for autonomous medical systems, focusing on enhancing visual perception and reasoning capabilities for radiology report generation. It features a multi-agent architecture that includes components for visual evidence extraction, knowledge graph construction, and a retrieval-guided drafting process, leading to significant improvements in benchmark metrics such as BLEU-1 (0.0493 to 0.3359) and ROUGE-L (0.0863 to 0.2440) on a chest radiograph dataset. This framework addresses the limitations of existing multimodal models by providing more reliable and interpretable diagnostic outputs, which is crucial for practitioners developing AI-driven healthcare solutions.