Abstract:To address the unreliable diagnostic performance of medical report generation methods for rare diseases and the insufficient medical contextual information, a diagnosis-guided report generation method integrating recursive classification and adaptive retrieval is proposed. Built upon visual features, a confidence-based recursive classifier is introduced to perform multi-round refinement on low-confidence samples, while Focal Loss is incorporated to enhance attention to rare categories. In addition, an adaptive retrieval is designed to dynamically select the most relevant historical reports based on visual features, providing supplementary medical context for report generation. Experimental results on the MIMIC-CXR dataset show that the proposed method achieves notable improvements in BLEU-n and ROUGE-L scores and demonstrates superior performance in diagnosis and description of rare diseases compared with existing methods such as UAR.