BAMT-Net: a boundary-aware multi-task framework for 3D breast ultrasound lesion segmentation and classification

Mammography and ultrasound have long been the primary modalities for breast cancer screening; however, both methods still have certain limitations. In mammography, high breast density can sometimes obscure the visualization of lesions [ 1 , 2 ]. While ultrasound is effective for displaying certain types of lesions, its heavy reliance on the operator’s experience can lead to diagnostic discrepancies. Unlike the standardized position and fixed parameters of mammography, handheld ultrasound scanning coverage, probe pressure, and gain adjustment vary from person to person, making complete standardization difficult. The image quality and diagnostic information acquisition are highly dependent on the operator’s scanning technique, instrument adjustment ability, and real-time interpretation skills [ 3 , 4 ].Furthermore, the two-dimensional images from traditional imaging techniques are often constrained by resolution and viewing angles, particularly when visualizing small or deep-seated lesions. Automated Breast Volume Scanner (ABVS), a novel three-dimensional breast imaging technology, is emerging as an increasingly popular choice [ 5 ]. Compared to traditional ultrasound, ABVS employs advanced 3D imaging methods to provide a comprehensive view of the internal breast structure, clearly presenting details of lesions, organ boundaries, and other abnormal areas [ 6 , 7 ]. During examinations, ABVS utilizes a wide-range high-frequency probe controlled by a robotic arm to automatically scan the breast along a preset path, eliminating the need for manual probe handling and movement by the operator [ 8 , 9 ]. This technology significantly reduces operator dependency and effectively prevents missed diagnoses caused by inexperience. This helps clinicians make more accurate diagnoses with more comprehensive and intuitive data, significantly improving screening accuracy and the reliability of early diagnosis [ 10 , 11 , 12 , 13 ].