Semi-Supervised Medical Image Segmentation using Swin Transformer with Focused Linear Self-Attention and Data Augmentation

Semi-supervised medical image segmentation methods have been proposed to address the issue of annotation scarcity in supervised learning by leveraging the information present in unlabeled data. However, current methods commonly rely on convolutional neural networks (CNNs) to build a unified model. They ensure the prediction consistency by applying minor perturbations to either the input or the model. This learning paradigm faces the following two primary challenges: (1) Reliable feature-level and pixel-level information are prone to being overlooked, making it difficult to learn the organizational structure and anatomical prior knowledge from a large amount of unlabeled data; (2) The inherent locality of CNN-based models hinders the global interaction of semantic information. Therefore, in this paper, we propose a novel semi-supervised segmentation model based on Swin Transformer with focused linear self-attention and data augmentation. Specifically, we firstly proposed a new data augmentation strategy based on multi-angle image transformation, which fully utilizes a large amount of unlabeled data to learn important information about organizational structures and organs from medical images. Secondly, we proposed the refined Swin Transformer based semi-supervised image segmentation model with focused linear self-attention in the Teacher and Student learning framework, which enables the model to learn global interaction information while concentrating the attention on important features, thereby achieving a higher performance. Finally, extensive experiments demonstrate that our approach outperforms state-of-the-art ones on two publicly available medical image segmentation benchmarks.