A progressive fine-tuning strategy for robust breast lesion detection across multiple mammography datasets

This study presents a YOLO (You Only Look Once)-based approach for the automated detection of breast lesions in mammography images; the corresponding implementation is provided in the Appendix. The proposed method adopts a continuous training scheme across three public mammography datasets, organized according to increasing annotation quality and domain complexity. Initially, the model is trained on the CBIS-DDSM dataset, which consists of digitized film mammograms with coarse annotations, achieving an mAP@0.5 of 45.6% and a precision of 50.7%. Subsequently, the resulting model is fine-tuned on the INBreast dataset, which comprises high-quality digital mammograms with expert-level annotations, leading to substantial performance improvements, with an mAP@0.5 of 76.73% and a precision of 83.32%. Finally, the refined model is further fine-tuned on the VinDr-Mammo dataset, yielding additional gains and reaching an mAP@0.5 of 80.55% and a precision of 80.07%. Overall, the results demonstrate consistent improvements in accuracy and robustness across heterogeneous mammographic datasets.