SEAS-Net: Segment Exchange Augmentation for Semi-supervised Brain Tumor Segmentation

Accurate segmentation of brain tumors is crucial for cancer diagnosis, treatment planning, and evaluation. However, semi-supervised brain tumor image segmentation methods often face the challenge of mismatched distribution between labeled and unlabeled data. This study proposes a Segment Exchange Augmentation for Semi-supervised Segmentation Network (SEAS-Net) based on a teacher-student model that exchanges labeled and unlabeled data segments. This approach fosters unlabeled data to grasp general semantics from labeled data, and the consistent learning for both types notably reduces the empirical distribution gap. The SEAS-Net consists of two stages: supervised pre-training and semi-supervised segmentation. In the pre-training stage, segment exchange strategies optimize labeled images. During the semi-supervised stage, the teacher network generates pseudo-labels for unlabeled images. Subsequently, labeled and unlabeled images through segment exchange and input to the student network for generating predictive segmentation templates. Pseudo-labels and real mixed supervised signals supervise these predictions. Additionally, the discriminator after the student network enhances the reliability of prediction results. This methodology excels on the BraTS2019 and BraTS2021 datasets, effectively mitigating data distribution disparities and substantially enhancing brain tumor segmentation accuracy.