Semi-supervised medical image segmentation method using multi-scale consistency adversarial learning

Medical image segmentation is an essential task in medical image analysis, playing a crucial role in improving the accuracy of disease diagnosis and the efficiency of treatment planning. Despite extensive research efforts and notable technological advancements of deep learning, its performance is often limited by the necessity for vast amounts of accurately labeled images, which are inherently costly and labor-intensive to procure. To mitigate this challenge, this study proposes a novel semi-supervised segmentation using Multi-Scale Consistency Adversarial Learning (MSCAL). By leveraging few annotated images, this method constructs a comprehensive data augmentation perturbation space, incorporating both image-level strong–weak perturbations alongside multi-scale feature perturbations. Furthermore, strong–weak consistency regularization and multi-scale adversarial learning strategies across diverse scales of the segmentation network are implemented. Furthermore, the method utilizes adaptive weighted pyramid consistency loss to encourage consistent predictions across scales, and emphasizes consistency in high-confidence regions through the confidence maps outputted by the discriminator. Finally, the advantages of our proposed model are rigorously evaluated on the ACDC and BraTS2019 datasets, where it is systematically compared against ten state-of-the-art semi-supervised methods. Experimental results demonstrate the model’s superiority across DSC, HD95, ASD metrics and $p$-value. Notably, with only 3 annotated samples, this method achieves at least 4.2%, 2.2%, and 1.6% gains in segmentation accuracy for the right ventricle, myocardium, and left ventricle, respectively. Ablation studies further corroborate this innovative framework enables the acquisition of richer feature representations and bolstering model robustness for semi-supervised medical image segmentation tasks.