Abstract
Ultrasound video-based breast lesion segmentation provides a valuableassistance in early breast lesion detection and treatment. However, existingworks mainly focus on lesion segmentation based on ultrasound breast imageswhich usually can not be adapted well to obtain desirable results on ultrasoundvideos. The main challenge for ultrasound video-based breast lesionsegmentation is how to exploit the lesion cues of both intra-frame andinter-frame simultaneously. To address this problem, we propose a novelSpatial-Temporal Progressive Fusion Network (STPFNet) for video based breastlesion segmentation problem. The main aspects of the proposed STPFNet arethreefold. First, we propose to adopt a unified network architecture to captureboth spatial dependences within each ultrasound frame and temporal correlationsbetween different frames together for ultrasound data representation. Second,we propose a new fusion module, termed Multi-Scale Feature Fusion (MSFF), tofuse spatial and temporal cues together for lesion detection. MSFF can help todetermine the boundary contour of lesion region to overcome the issue of lesionboundary blurring. Third, we propose to exploit the segmentation result ofprevious frame as the prior knowledge to suppress the noisy background andlearn more robust representation. In particular, we introduce a new publiclyavailable ultrasound video breast lesion segmentation dataset, termed UVBLS200,which is specifically dedicated to breast lesion segmentation. It contains 200videos, including 80 videos of benign lesions and 120 videos of malignantlesions. Experiments on the proposed dataset demonstrate that the proposedSTPFNet achieves better breast lesion detection performance thanstate-of-the-art methods.