A Multi-Task Local-Global U-Net for Simultaneous Segmentation and Classification of Breast Lesions in Ultrasound Images

Abstract

Breast cancer remains the most frequently diagnosed cancer among women worldwide and a leading cause of cancer-related death, with an estimated 2.3 million new cases and more than 670, 000 deaths worldwide in 2022. Breast ultrasonography is widely used as an adjunct diagnostic modality because it is non-ionizing, relatively affordable, and suitable for examining dense breast tissue. Nevertheless, speckle noise, low contrast, and indistinct lesion boundaries continue to challenge automated lesion segmentation and diagnostic classification. This study presents LGU-Net (Local-Global U-Net), a parameter-conscious multi-task architecture that simultaneously performs binary lesion segmentation and three-class image classification comprising normal, benign, and malignant categories. The architecture combines a depthwise-pointwise convolutional encoder, a two-stage sparse-to-dense local-global attention bottleneck, an ASPP-Lite context module, and a U-Net-style decoder with concurrent spatial and channel squeeze-and-excitation attention. Segmentation and classification are trained end to end using a combined Dice, focal, and Intersection over Union segmentation objective together with uncertainty-based task balancing. On the Breast Ultrasound Images dataset, LGU-Net achieved a mean Dice score of 0.7925 and a mean IoU of 0.7228 for segmentation, together with a classification accuracy of 0.8333 and a macro-F1 score of ~0.83. These results demonstrate the feasibility of jointly modeling lesion localization and three-class diagnostic prediction within a compact shared architecture.