Evaluation of Deep Learning Techniques for Automatic Lesion Segmentation in Mammography Images

The increasing breast cancer incidence and importance on early-stage diagnoses drives advances in analysis of screening imaging modalities, assisting efficient patient prioritization and decision support. This work addresses breast lesion segmentation problem in mammograms collected within INCISIVE project from multiple hospital centers, aiming to assess the realistic performance on clinical-grade image quality with minimal data curation. Global-Local Activation Maps (GLAM) as weakly supervised localization approach and medical transformer (MT) based models were evaluated on the task. We further propose segment level performance metric as better aligned to human reader evaluation, to complement commonly used pixel-level evaluation. The results indicate that supervised approaches achieve better performance on pixel-level metrics, with F1-score of 0.487, yet both approaches are well aligned in segment-level metrics with 0.664 F1-score for MT model and 0.644 for GLAM. We further identify main difficulties in mammogram analysis and possible alleys for further improvements to achieve robust and re producible performance.