Image Pseudo Label Consistency Exploitation for Semi-supervised Pathological Tissue Segmentation

Supervised deep learning-based segmentation methods help doctors to identify regions of human tissues and lesions on pathological images and diagnosis diseases. However, due to the huge sizes of pathological images and the fragile shapes of human tissues and lesions, labeling large scale training data for the supervised deep learning methods is prohibitive. Semi-supervised learning methods generate pseudo-labels of unlabeled data and utilize the information from both labeled and unlabeled data to reduce the required amount of labeled data for training. One of the critical issues of semi-supervised learning is to generate consistent pseudo-labels for similar samples. To improve the consistency of the pseudo-labels, we propose an image pseudo label consistency exploitation method to regularize the models to generate similar predictions for similar samples by considering the image consistent loss and set consistent loss with the help of data augmentations of the unlabeled images. The experiments on two pathological segmentation datasets show the superior of the proposed method over state-of-the-art methods.