Patch Shuffle and Pixel Contrast: Dual Consistency Learning for Semi-supervised Lung Tumor Segmentation

Semi-supervised learning is a promising approach for medical image segmentation with limited labeled data. Though existing consistency learning based SSL methods achieve convincing results, they neglect the finer-grained information. In this work, we propose a novel dual consistency learning (DCL) method based on characteristics of medical images for semi-supervised lung tumor segmentation. For patch shuffle consistency learning, image patches are shuffled as a strong-augmented view to improve both the student and teacher models in the mean teacher framework. For pixel contrast consistency learning, we construct a memory bank by high-quality pixel features updated with reservation and obtain anchors from wrongly classified tumor and high-confident background features, making the pixel-level feature space more discriminative. Experiments on three lung tumor datasets demonstrate the effectiveness of our method for semi-supervised medical image segmentation.