LiteBUS-SSM: a boundary-preserving lightweight state-space U-Net for breast ultrasound lesion segmentation

Abstract

Breast ultrasound lesion segmentation is challenged by speckle noise, weak boundaries, acoustic shadows, heterogeneous morphology, and false positives in normal images. This paper proposes LiteBUS-SSM, a lightweight U-shaped network for boundary-preserving and parameter-efficient breast ultrasound segmentation. BP-LKA is placed only in shallow encoder stages to enhance contour and texture cues, while a single bottleneck BiVSS-Calibrator models global context at low spatial resolution. Training combines deep-supervised Tversky optimization with boundary-weighted log its distillation, and adaptive small-component filtering suppresses tiny isolated responses during inference. Experiments on patient-level BUS-UCLM validation and BUSI cross-dataset assessment report Dice, IoU, HD95, N-FPR, statistical comparison, and deployment cost. LiteBUS-SSM achieves a favorable accuracy–efficiency balance, with its advantage mainly reflected in lower parameter cost, lower boundary error, and false-positive control rather than a statistically decisive Dice improvement over the strongest baseline.