AstMatch: Adversarial self-training consistency framework for semi-supervised medical image segmentation

Abstract

Semi-supervised learning (SSL) has demonstrated significant potential in medical image segmentation, primarily through consistency regularization and pseudo-labeling. However, many SSL approaches focus predominantly on low-level consistency, neglecting the significance of pseudo-label reliability. Therefore, in this work, we propose an adversarial self-training consistency framework (AstMatch). First, we design an adversarial consistency regularization (ACR) approach to enhance knowledge transfer and strengthen prediction consistency under varying intensity perturbations. Second, we incorporate a feature matching loss within adversarial training to achieve high-level consistency regularization. Furthermore, we present the pyramid channel attention (PCA) and efficient channel and spatial attention (ECSA) modules to enhance the discriminator’s effectiveness. Finally, we propose an adaptive self-training (AST) approach to ensure high-quality pseudo-labels. The proposed AstMatch has been extensively evaluated with state-of-the-art SSL methods on three publicly available datasets. Experimental results across different labeled ratios demonstrate that AstMatch outperforms other existing methods, achieving new state-of-the-art performance. Our code is publicly available at http://github.com/GuanghaoZhu663/AstMatch.