Multi-task Semi-supervised 3D Medical Image Segmentation Based on CNN and self-attention

Deep learning techniques have recently achieved remarkable results in medical image analysis. However, these methods typically need full supervision and demand extensive labeled datasets, which are laborious and hard to get. Semi-supervised methods that leverage unlabeled data in model training have become a promising area of research. We propose that integrating reconstruction tasks with segmentation tasks during model training can enable the model to acquire more comprehensive semantic information from images. Specifically, we propose MTSS-Net, a framework based on semi-supervised learning and multi-task learning, designed for 3D image segmentation. Our model jointly learns 3D image reconstruction and segmentation tasks. Additionally, our segmentation network integrates convolutional neural networks (CNNs) with self-attention mechanisms to improve the processing of local information and the modeling of global context. During training, we followed the mean teacher strategy, which enforces consistency between model outputs to facilitate the learning of more robust features. Experimental results on a 3D medical image segmentation dataset demonstrate the feasibility and superiority of MTSS-Net compared to other methods.