Integrating Clinically-Aligned Chain-of-Thought Reasoning for Radiology Report Generation

Automated radiology report generation holds significant potential to improve diagnostic accuracy and accelerate clinical workflows. However, current methods fail to capture nuanced reasoning patterns of radiologists due to their limited modeling of the complex diagnostic logic. To this end, we propose integrating Clinically-Aligned Chain-of-Thought (CACoT) reasoning pathways into the report generation process. Our method first establishes a hierarchical diagnostic schema organized by imaging projections and anatomical structures, thereby ensuring a structured, interpretable workflow. Within each anatomical structure, the diagnostic reasoning process is decomposed into a five-stage pipeline encompassing region localization, disease recognition, severity assessment, possibility reasoning, and evidence-based interpretation. These modular reasoning chains are subsequently synthesized across anatomical structures to form a comprehensive report. Recognizing the substantial cost of Chain-of-Thought annotations and the heterogeneous value of available training data, we further integrate an active learning paradigm to prioritize the annotation of highly informative samples, guided by uncertainty and inconsistency estimation. Extensive experiments on benchmark datasets, evaluated with specialized medical metrics, demonstrate that our CACoT effectively captures the clinical reasoning patterns and achieves excellent performance.