Semi-Supervised Semantic Segmentation with Scope Consistency Constraint in TRUS Image

Abstract

In full supervised learning, the number of pixel-level labels for learning is small, and the complexity of neural network learning objectives is complex. A cross-label training framework based on Scope Consistency Constraint (SCC) is proposed for this challenge. Our framework mainly establishes task-level consistency through the outputs of two backbone networks and provides each other with pseudo labels for each other to learn. In addition, our framework introduces weak labels into the learning process of the segmentation network, significantly reducing the labeling cost. The experiment shows that the segmentation accuracy is better than the original segmentation network framework with the same amount of data. In addition, the semi-label data is introduced. The framework presented in this paper is not limited to a specific segmentation framework.