Ultrasound Image Segmentation via a Multi-scale Salient Network

Deep learning methods have significantly advanced the field of ultrasound imaging analysis. However, the low quality of ultrasound images remains a bottleneck, especially in the case of tumor segmentation. Motivated by these facts, this paper introduces a novel method called the Deep Saliency Attention Network (DSA-net) to segment breast tumors in ultrasound images. The DSA-net is developed to learn multi-scale characteristics of tumors in a noisy environment and with limited data. The proposed method achieves accurate segmentation performance by combining saliency and attention mechanisms. The effectiveness of DSA-net is evaluated on three benchmark databases of breast ultrasound images, benchmarked against various widely used and state of-the-art (SOTA) methods (e.g. U-net, AAU-net, U2-net, and Attention U-net). The proposed DSA-net method demonstrates SOTA visual and quantitative improvements evaluated on the breast cancer databases.