Lightweight U-Net for Breast Ultrasound Image Segmentation

Purpose: Accurate segmentation of breast lesions in ultrasound images is essential for early breast cancer diagnosis and treatment planning. This work introduces a lightweight U-Net architecture optimized for clinical deployment. Methods: We replace standard U-Net encoder blocks with a MobileNetV2 backbone trained from scratch, yielding 3.10 M parameters and 0.72 GFLOPs. Evaluations use the BUS-BRA dataset. Results: On the BUS-BRA dataset, the proposed model achieves a Dice coefficient of 88.88% and a mean IoU of 81.26%, surpassing the standard U-Net, Attention U-Net, and self-supervised baselines, while maintaining an efficient inference time of 1.53 ± 0.33 s per image. Conclusion: Combining a lightweight encoder with attention mechanisms delivers high segmentation accuracy and computational efficiency, making it well-suited for breast ultrasound applications on resource-limited hardware.