Dual Attention Based Uncertainty-aware Mean Teacher Model for Semi-supervised Cardiac Image Segmentation

Recently, many fully supervised deep learning based methods have been proposed for automatic cardiac segmentation. However, it is very expensive and time-consuming to annotate data for the task. In this paper, we present a novel dual attention based uncertainty-aware mean teacher semi-supervised framework (DA-UAMT) for cardiac image segmentation. The framework consists of a teacher model and a student model with the same structure and the student model learns from the teacher model by minimizing a segmentation loss generated from labeled images and a consistency loss generated from unlabeled images with respect to the targets of the teacher model. To enable the student model learn from more reliable targets, we introduce the Monte Carlo Dropout which estimates target uncertainty, and a novel dual attention mechanism which helps the network to focus on information in shape and channel dimension. To evaluate the proposed method, we conducted experiments on MICCAI 2017 Automated Cardiac Diagnosis Challenge (ACDC) dataset. Experiments show that our proposed DA-UAMT model is effective in utilizing unlabeled data to obtain considerably better segmentation of cardiac.