Semi-Supervised Medical Image Segmentation via Multi-Group Network with Consistency Among Heterogeneous Loss Supervision

Semi-Supervised Learning (SSL) is appealing for reducing the annotation cost for training medical image segmentation models. State-of-the-art SSL methods typically generate multiple predictions via multiple networks/branches or perturbed inputs, then employ cross-supervision or consistency regularization to leverage unannotated images. Nevertheless, these approaches are often encumbered by either high computational costs or excessive time consumption due to their complexity. To solve these problems, we introduce an efficient framework Multi-Group network with Consistency among Heterogeneous Loss (MGCHL) for SSL. First, a Multi-Group Convolutional Network (MGCNet) is proposed to efficiently produce multiple predictions, each from an independent group of features. Furthermore, to encourage inter-group interaction for robust learning, we propose heterogeneous supervision where each group is supervised by a distinct loss on labeled images, and then inter-group consistency is imposed on unlabeled images for regularization. Experimental results demonstrated that our method outperforms four state-of-the-art methods on both skin lesion segmentation in the ISIC 2018 dataset and heart structure segmentation in the ACDC dataset.