Local augmentation based consistency learning for semi-supervised pathology image classification

Abstract

Background and objectiveLabeling pathology images is often costly and time-consuming, which is quite detrimental for supervised pathology image classification that relies heavily on sufficient labeled data during training. Exploring semi-supervised methods based on image augmentation and consistency regularization may effectively alleviate this problem. Nevertheless, traditional image-based augmentation (e.g., flip) produces only a single enhancement to an image, whereas combining multiple image sources may mix unimportant image regions resulting in poor performance. In addition, the regularization losses used in these augmentation approaches typically enforce the consistency of image level predictions, and meanwhile simply require each prediction of augmented image to be consistent bilaterally, which may force pathology image features with better predictions to be wrongly aligned towards the features with worse predictions. MethodsTo tackle these problems, we propose a novel semi-supervised method called Semi-LAC for pathology image classification. Specifically, we first present local augmentation technique to randomly apply different augmentations produces to each local pathology patch, which can boost the diversity of pathology image and avoid mixing unimportant regions in other images. Moreover, we further propose the directional consistency loss to enforce restrictions on the consistency of both features and prediction results, thus improving the ability of the network to obtain robust representations and achieve accurate predictions. ResultsThe proposed method is evaluated on Bioimaging2015 and BACH datasets, and the extensive experiments show the superior performance of our Semi-LAC compared with state-of-the-art methods for pathology image classification. ConclusionsWe conclude that using the Semi-LAC method can effectively reduce the cost for annotating pathology images, and enhance the ability of classification networks to represent pathology images by using local augmentation techniques and directional consistency loss.