Weakly-Supervised Ultrasound Video Segmentation with Minimal Annotations

Ultrasound segmentation models provide powerful tools for the diagnosis process of ultrasound examinations. However, developing such models for ultrasound videos requires densely annotated segmentation masks of all frames in a dataset, which is unpractical and unaffordable. Therefore, we propose a weakly-supervised learning (WSL) approach to accomplish the goal of video-based ultrasound segmentation. By only annotating the location of the start and end frames of the lesions, we obtain frame-level binary labels for WSL. We design Video Co-Attention Network to learn the correspondence between frames, where CAM and co-CAM will be obtained to perform lesion localization. Moreover, we find that the essential factor to the success of extracting video-level information is applying our proposed consistency regularization between CAM and co-CAM. Our method achieves an mIoU score of 45.43% in the breast ultrasound dataset, which significantly outperforms the baseline methods. The codes of our models will be released.