MammoExpert: Benchmarking Chain-of-Thought Reasoning in Mammography Diagnosis

MammoExpert is introduced as a novel mammography dataset featuring Chain-of-Thought (CoT) reasoning annotations across three diagnostic phases, comprising 2,379 images and 42 radiographic features annotated by nine senior radiologists. The dataset enhances breast lesion classification accuracy, achieving a 7.1% improvement when combined with the CBIS-DDSM dataset and an additional 4% gain through CoT reasoning training on the MammoExpert test set, with similar improvements noted on INBreast and Vindr datasets. This benchmark is significant for practitioners as it facilitates the development of interpretable AI models in mammography diagnosis, emphasizing structured reasoning in clinical decision-making.