Knowledge-Guided and Reinforced Selective State Space Model for radiology report generation.

Abstract

To address the limitations of "black box" sequence models and simplistic reward functions in radiology report generation, we aim to improve clinical accuracy and linguistic quality by integrating explicit medical knowledge and a clinically-aligned reinforcement learning strategy. We propose a novel framework, the Knowledge-Guided and Reinforced Selective State Space Model (KGR-SSM). This model synergistically integrates three key components: (1) a medical knowledge graph to explicitly guide the semantic understanding of visual features; (2) an efficient Mamba-based encoder for processing high-resolution images; and (3) a hybrid reward function for reinforcement learning that optimizes for both NLG metrics and clinical accuracy. Extensive experiments on the public IU X-ray and MIMIC-CXR datasets demonstrate that our proposed KGR-SSM achieves state-of-the-art performance. The model significantly outperforms existing methods across a comprehensive suite of evaluation metrics, including both linguistic and clinical efficacy measures. The integration of structured medical knowledge and a clinically-oriented hybrid reward function effectively enhances the accuracy and reliability of automated radiology report generation. The KGR-SSM framework provides a robust and promising solution for this critical clinical task, bridging the gap between technical performance and clinical utility.