CellMix: A General Instance Relationship-Based Method for Data Augmentation Toward Pathology Image Classification

In the pathology image analysis, obtaining and maintaining high-quality annotated samples is an extremely labor-intensive task. To overcome this challenge, mixing-based methods have introduced new relationships to traditional preprocessing data augmentation techniques. Nonetheless, these methods fail to fully consider the unique features of pathology images, such as local specificity, global distribution, and inner/outer sample instance relationships. To better comprehend these characteristics and create valuable pseudosamples, we propose the CellMix framework, which employs a novel distribution-oriented in-place shuffle approach. The images are divided into patches based on the granularity of pathology instances, and the patches are in-place shuffled within the same batch. Thus, the locational relationships among instances can be effectively preserved while new relationships can be further introduced. Moreover, inspired by curriculum learning (CL), a loss-driven strategy is designed to control the relationship augmentation. This strategy enables the model to adaptively explore the instances at multiple scales and efficiently handle distribution-related noise under various difficulties. Our experiments in pathology image classification tasks demonstrate state-of-the-art (SOTA) performance on seven distinct datasets. This innovative instance relationship-centered method sheds light on general data augmentation for pathology image classification. The associated codes are available at: https://github.com/sagizty/CellMix.