Enhancing Semi-Supervised Brain Tumor Segmentation via Dual-Uncertainty Guided 2.5-D CNN–Transformer and Voxel-Wise Contrastive Learning

Data annotation in medical image segmentation is time-consuming and expensive. Semi-supervised learning presents a viable solution. However, unlike organ segmentation, current semi-supervised methods often produce highly uncertain and erroneous predictions in the boundary regions of brain tumors due to their variable shapes, uneven distribution, and complex boundaries. In this paper, we propose a 2.5D CNN-Transformer semi-supervised segmentation model (SemiCTrans) that incorporates 3D spatial information while leveraging the strengths of different network architectures to capture local and global image features. Additionally, we design a dual-uncertainty guided module (DUGM) that automatically retains reliable prediction regions in unlabeled data. Within these reliable regions, we enforce a consistency constraint on the predictions of the CNN-Transformer network and conduct voxel-level intra-and inter-class contrastive learning for tumor segmentation. Extensive experiments on two datasets, including one publicly available dataset (BraTS2020 dataset) and one private dataset, demonstrate that the proposed SemiCTrans achieves higher performance and superior robustness than state-of-the-art brain tumor image segmentation methods. Remarkably, even with just 1% labeled data, our method achieves a Dice similarity coefficient (DSC) score of 74.7% on the BraTS2020 dataset. Code is available at https://github.com/zydlsd/SemiCTrans.