An Automatic Ultrasonic Segmentation Method by Two-stage Semi-supervised Learning Strategy

Medical image segmentation plays an important role in computer aided diagnosis. This requires a lot of labeling, and the cost is significant. As a result, a new deep learning model has been proposed to improve the workload of radiologists' labeling. This method contains two stage: 1) Supervised learning network to frame the region of interest to narrow the segmentation scope for the next step; 2) Unsupervised learning is used to accurately segment the prostate. Especially, only in the first part is a rough notation needed. In addition, we use two modalities containing B-mode and elastography to validate the segmentation network and adopt multi-center data to guarantee the robustness of the model. In the end, the network is compared with other advanced semi-supervised DL methods such as DAN and ASDNet according to several quantitative indicators, and achieve the state-of-the-art performance.