Dual Diversity and Pseudo-Label Correction Learning for Semi-Supervised Medical Image Segmentation

Abstract

Semi-supervised medical image segmentation has recently gained increasing research attention as it can reduce the need for large-scale annotated data. Current mainstream methods usually adopt two sub-networks and encourage the two models to make consistent predictions for the same segmentation task through consistency regularization. However, the scarcity of medical samples reduces the effectiveness of consistency constraints, and this problem may be further exacerbated by the influence of noisy pseudo-labels. In this work, we propose a novel co-training framework based on dual diversity and pseudo-label correction learning (DDPCL) to address these challenges. Specifically, firstly, we design a dual diversity learning strategy, in which data diversity fully mines the potential information of limited training samples through the CutMix operation, and feature diversity promotes the model to learn complementary feature representations by minimizing the similarity between the features extracted by the two sub-networks. Secondly, we propose a pseudo-label correction learning strategy, which regards the inconsistent region where the pseudo-labels predicted by the two sub-networks are different as potential bias regions, and guides the models to correct the bias in these regions. Extensive experiments on three public datasets (ACDC, LA and Pancreas-NIH datasets) validate that the proposed method outperforms the state-of-the-art semi-supervised medical image segmentation. The code is available at http://github.com/ddd0420/ddpcl.