Semi-supervised Segmentation of Medical Images

This article is devoted to the development of a method (algorithm) of medical image segmentation based on semi-supervised learning. Semi-supervised learning methods are shown to have significant potential for improving medical image segmentation through effective use of unlabeled data. However, challenges remain in adapting these methods to the specific characteristics of medical images, such as high variability, class imbalance, and the presence of noise and artifacts. To overcome these difficulties, it is proposed to integrate several approaches (consistency regularization, pseudo-labeling, average teacher model) into a single structure. To increase the robustness and generalizability of the model for different imaging methods, we include industry-specific data supplements tailored to the unique characteristics and challenges of each method. Large-scale experiments on magnetic resonance imaging, computed tomography, and optical coherence tomography datasets demonstrate that the proposed framework significantly outperforms fully supervised and individual semisupervised learning methods, especially in scenarios with low data labeling.