LBA-Net: Lightweight Boundary-Aware Network for Robust Breast Ultrasound Image Segmentation

Accurate yet efficient segmentation of breast ultrasound (BUS) images is essential for early tumour diagnosis, but is still hampered by speckle noise, low-contrast and ill-defined lesion boundaries. Although deep learning models have improved accuracy, their prohibitive complexity limits deployment on portable or edge devices. We propose LBA-Net, a lightweight boundary-aware network that delivers competitive accuracy with minimal computational cost. Built upon a MobileNetV3-Small encoder, LBA-Net employs a multi-scale ASPP module and a novel LBA-Block that fuses efficient channel and spatial attention to enhance tumour features while suppressing noise. A dual-head decoder, supervised by a boundary-sensitive composite loss, further refines contour precision. On the public BUSI(Breast Ultrasound Images) dataset (780 lesion images with masks), five-fold cross-validation yields a mean Dice of 93.36% and IoU of 88.76%—among the highest results reported for models under 5 GFLOPs. The network contains only 1.98 M parameters and 3.34 GFLOPs per inference, @512×512 input. Real-time benchmarking on an NVIDIA A100 shows 122 FPS (FP32, batch=1) and 779 FPS (torch.compile + FP16, batch=1), scaling to 1943 samples s⁻¹ at batch=4; even on Intel Xeon CPU (single-thread) it reaches 8.6 FPS, enabling point-of-care deployment.