Semi-supervised segmentation on medical images with pseudo label calibration and neural process

Pseudo supervision has demonstrated empirical success in semi-supervised segmentation tasks by effectively leveraging unlabeled data, but it unavoidably encounters the problem caused by noisy pseudo labels. Existing methods against noisy pseudo labels heavily rely on the model confidence which may be unreliable in some cases. In this work, we mitigate the impact of noisy pseudo label from two aspects. First, we propose a new pseudo supervision with dual confidence that can weaken the supervision from the noisy pseudo labels with high-level confidence. This mechanism can make the segmentation model more robust to the unreliable confidence. Second, to improve the quality of pseudo labels, we propose a novel regularization method for calibrating the noisy pseudo labels. Different from existing works, this regularization is independent of the model confidence, it focuses on the distribution consistency between labeled and pseudo-labeled data. To this end, a joint optimization based on the theory of neural process is introduced into the pseudo label calibration. We verify the effectiveness of the proposed method on the medical image (MI) segmentation that is important for the disease treatment. It experimentally shows that our method can bring a practical improvement to the MI segmentation task in comparison with the existing state-of-the-art semi-supervised learning methods. Our method provides some new insight into the pseudo supervision.