10 Abnormality-Focused Two-Phase Deep Learning for Mammographic Lesion Segmentation: Strategies and Limitations

Background

Mammography reduces breast cancer mortality through early detection; however, radiologists miss up to 30% of cancers, largely due to perceptual limitations and interpretive variability. Deep learning–based segmentation models have shown promise in mammographic lesion detection, but performance is limited by severe class imbalance, small lesion size, and loss of spatial detail during image downsampling. Many existing models are trained on datasets dominated by normal examinations, restricting their ability to learn nuanced abnormal features. This study aimed to evaluate whether a 2-phase, abnormality-focused deep learning approach could improve lesion segmentation while characterizing key technical limitations inherent to mammography.

Materials and Methods

A retrospective segmentation study was conducted using annotated abnormal mammograms from a public dataset. Only images with radiographically confirmed lesions and ground-truth masks were included. A customized U-Net architecture was trained using a 2-phase strategy: (1) lesion-centered, patch-based pretraining to emphasize abnormal features, followed by; (2) full-image fine-tuning with hybrid sampling to preserve small-lesion sensitivity. Training employed focal loss and Tversky loss to address class imbalance. Performance was evaluated on an independent test set using Dice similarity coefficient, intersection-over-union (IOU), pixel accuracy, Hausdorff distance, and lesion-level detection rate (IOU >0.10). Size-stratified analyses were performed.

Results

On the test set (n = 55), the model achieved a mean Dice coefficient of 0.58 and detected 77.8% of lesions. Detection rates increased with lesion size: 73.0% for small (<500 pixels), 84.6% for medium, and 100% for large lesions. Among detected lesions, segmentation accuracy was high for larger abnormalities (mean Dice, 0.91). Missed detections predominantly involved very small lesions approaching the resolution limit after downsampling.

Conclusions

A 2-phase, abnormality-focused training strategy enables robust segmentation of moderate-to-large mammographic lesions but remains constrained by resolution-dependent failure in small abnormalities. These findings highlight both the promise of abnormal-focused deep learning and the persistent technical challenges of mammographic lesion segmentation, underscoring the need for multi-scale architectures and higher-resolution inference before clinical deployment.