The Semi-supervised Adversarial Self-integration Network for Medical Image Segmentation

To overcome the limitations of the traditional U-Net architectures, such as the ineffective context extraction and restricted receptive fields, a novel semi-supervised adversarial self-integration network that consists of segmentation and discriminative network, is proposed. The former used the semi-supervised learning strategy by integrating the convolutional neural networks and Vision Transformers, while the latter uses the adversarial consistency training with two consistency-based discriminators to capture the prior relationships. Because of the attention-based dynamic convolution, the network weights can be adaptively adjusted according to the sample’s structure, thereby the feature representation was improved and the overfitting risks was reduced. By comparing five methods on the ACDC, LA, and Pancreas datasets, experimental results demonstrated that the Dice coefficient, Jaccard coefficient, Hausdorff distance and the average surface distance was respectively improved by 3.4%~3.9%, 2.9%~4.0%, 43.5%~53.4%, 65.1%~68.7%, and better segmentation results can be also obtained for the case of fewer labeled data.