Multi-angle Dual-task Consistency for Semi-Supervised Medical Image Segmentation

Abstract

Semi-supervised learning has shown its effectiveness in the field of medical image segmentation tasks, since it mitigates the heavy burden of obtaining abundant reliable annotations, which is scarce and time-consuming, by leveraging the available unlabeled data. In this paper, we propose a novel Multi-angle Dual-task Consistency Network (MDC-Net) for semi-supervised medical image segmentation. Particularly, our MDC-Net can be formulated as a model similar to the Mean Teacher with both employing exponential moving average (EMA). The difference is that we utilize one of the networks to augment unlabeled data, thus generating its pseudo labels and exploiting MixUp on both labeled and unlabeled data, whose outputs are then utilized as the inputs of another dual-task network based on region-based shape constraints and boundary-based surface mismatch. Finally, by jointly learning pseudo labels, the segmentation probability map and the signed distance map of the target, our framework not only overcomes the problem of incomplete segmentation caused by noise in pseudo labels, but also enforces geometric constraints and makes the model more robust. Experimental results on the public Left Atrium (LA) database show that our method achieves performance gains by incorporating unlabeled data and outperforms six state-of-the-art semi-supervised segmentation methods.