Comparative Analysis of Attention Mechanisms in U-Net Variants for Breast Cancer Image Segmentation

Accurate segmentation of breast cancer images is crucial for effective detection and treatment. This work evaluates the efficacy of various Attention U-Net models with different attention mechanisms, including SE (Squeeze-and-Excitation) blocks and Swish activation functions, for breast cancer image segmentation. We assessed these models using several metrics: IoU, precision, recall, F1 score, MAE, dice coefficient, and specificity. The SE-Swish U-Net variant achieved the highest performance metrics, with a mean IoU of 77.71%, precision of 78.6%, recall of 73.5%, F1 score of 75.9%, dice coefficient of 75.91%, and specificity of 96.70%. The SE-Swish model also demonstrated superior accuracy and efficiency compared to other variants, though it required longer training and execution times. Qualitative analysis supported the SE-Swish model’s ability to handle complex lesion borders and low-contrast areas. This work highlights how advanced attention mechanisms in U-Net architectures can enhance breast cancer image segmentation.