Reinforcement Learning for Improved Alignment in Radiology Reports Generation

Chest X-rays are crucial for diagnosing various medical conditions, but generating accurate reports from them is challenging due to their complexity. Traditional models often miss critical anomalies and produce vague content, overlooking specific clinical signs. To address these issues, we introduce a reinforcement learning framework that improves alignment between radiology images and their corresponding reports. The framework leverages high-level concepts and uses a two-stage generation process combined with reinforcement learning fine-tuning to enhance the analysis of complex medical image patterns, achieving precise text-image alignment. Evaluations on the IU X-Ray and MIMIC-CXR datasets demonstrate our model's superior performance. Further analysis shows significant improvement in aligning information between X-rays images and textual reports, leading to more relevant medical reports.