Dual Consistency Regularization for Semi-supervised Medical Image Segmentation

Abstract

Recently, deep neural networks trained with limited amount of labeled data often yield uncertain predictions of pattern recognition tasks. Especially for medical image segmentation, existing methods are prone to produce inaccurate predictions in target edge regions due to the low contrast at organ boundaries. To address the above problem, we propose a novel dual consistency regularization network (DC-Net) for semi-supervised medical image segmentation, which can obtain low-entropy decision boundaries by performing consistency predictions under model-level and task-level perturbations. Specifically, our network comprises a shared encoder and multiple decoders with different up-sampling strategies. Each decoder is equipped with two branches for dual-task output. For model consistency, the cross-consistency loss is designed between the segmentation map and the pseudo-labels across different models, aiming to minimize the discrepancy among different model outputs. For task consistency, we promote consistency between the segmentation maps and the pixel-level probability maps transformation from the signed distance maps (SDM), thereby constructing the geometric contours of the target to achieve more precise segmentation boundaries. Experimental results on the public Left Atrium dataset have shown that DC-Net achieved Dice scores of 91.29% and 89.75% with 20% and 10% labeled data respectively, which surpasses the other six current promising methods.