Semi-supervised Medical Image Segmentation with Confidence Calibration

The lack of high-quality expert labeled data is a common shortfall for medical image segmentation, which promotes semi-supervised learning scheme to an active research topic. The pseudo-labeling technique has been demonstrated to be a powerful module in semi-supervised segmentation framework for leveraging unlabeled data. However, simple generated pseudo labels are inevitably noisy and limited by the introduced confirmation biases, for the reason that the prediction errors of these pseudo labels would enhance the misleading to the segmentation network. In this paper, we propose to estimate the prediction confidence during the training process and further utilize the confidence to calibrate the pseudo label with the purpose to mitigate the confirmation bias problem. To emphasize, the pixel-wise confidence of the prediction results are learned through an adversarial network, while untrustworthy areas could be determined based on the prediction confidence. Rectified pseudo labels on untrustworthy areas are modified and further be utilized for medical image segmentation. The effectiveness on segmentation performance and noisy pseudo label calibration are proved by comparing several supervised or semi-supervised methods on BraTS2015 dataset and other two 3D medical image datasets.