Hierarchical Augmentation Consistency Learning for Semi-Supervised Medical Image Segmentation

While convolutional neural network (CNN) based methods have driven pivotal advancements in medical image segmentation, they remain constrained by the heavy reliance on large-scale labeled datasets. Semi-supervised learning(SSL) has recently emerged as a promising alternative to mitigate this limitation by effectively leveraging unlabeled data. In this paper, we propose a novel semi-supervised medical image segmentation method based on a teacher-student network. We perform strong and weak data augmentations on the original unlabeled data to broaden the data distribution while ensuring predictive consistency between augmented variants. Specifically, the teacher model generates stable pseudo-labels in weakly augmented images, while the student model learns how to generate corresponding predictions in strongly augmented data. Moreover, the weakly augmented images and labeled images are generated as mixed images by a copy-paste strategy, while the predictions of the mixed images on the student model are constrained for consistency regularization by the associated outputs in the same copy-paste way. We conduct extensive experiments on the “Fetal Ultrasound Grand Challenge: Semi-Supervised Cervical Segmentation” dataset for the segmentation of anterior lip and posterior lip, and the results show that our method outperforms currently advanced semi-supervised image segmentation methods.