Dual consistency semi-supervised learning for 3D medical image segmentation

Abstract

Semi-supervised learning (SSL) methods can utilize a large amount of unlabeled data to alleviate the problem of scarce labeled data. Consistency regularization as a classic SSL strategy, has achieved significant results in this field. However, due to the low contrast of organ contours in medical images, existing approaches are prone to uncertain predictions at the segmentation boundaries and the hard-to-identify blurry areas. To address these issues, we propose a dual consistency semi-supervised medical image segmentation network (DC-Net). Specifically, it can reduce the prediction uncertainty to obtain low-entropy decision boundaries by performing consistency prediction under model-level and task-level perturbations. We first design a cross-consistency loss between the segmentation map and the pseudo-labels across different models, aiming to encourage the models to maintain consistent predictions in blurry areas. Then, consistency constraints are applied between the segmentation task and the geometric perception task to construct the geometric contours of the target, thereby obtaining more precise boundary distance information. The experiments on four public medical image datasets (including LA, Pancreas CT, ACDC and PROMISE2012) demonstrate that DC-Net has achieved the state-of-the-art performance over other advanced semi-supervised methods.