Confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation

The unsupervised cross-modality image segmentation has gained much attention. Many methods attempt to align different modalities via adversarial learning. Recently, self-training with pseudo labels for the unsupervised target modality has also been widely used and achieved very promising results. The pseudo labels are usually obtained by selecting reliable predictions whose highest predicted probability is larger than an empirically set value. Such pseudo label generation inevitably has noise and training a segmentation model using incorrect pseudo labels could yield nontrivial errors for the target modality. In this paper, we propose a confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation. Specifically, we independently initialize two networks with the same architecture, and propose a novel confidence-weighted Dice loss to mutually supervise the two networks using their predicted results for unlabeled data. In this way, we make full use of all predictions of unlabeled images and leverage the prediction confidence to alleviate the negative impact of noisy pseudo labels. Extensive experiments on three widely-used unsupervised cross-modality image segmentation datasets (i.e., MM-WHS 2017, Brats 2018, and Multi-organ segmentation) demonstrate that the proposed method achieves superior performance to some state-of-the-art methods.