MH-CLR: A Self-Supervised Learning Strategy Based on Hard Sample Mining and Masked Autoencoders

The integration of Whole Slide Imaging (WSI) technology with artificial intelligence has significantly enhanced the analysis and application of histopathological slides. However, existing supervised learning methods for WSI analysis demand extensive time and effort to annotate high-resolution slide images. In contrast, Self-Supervised Learning (SSL) pretraining strategies, which do not depend on explicit data annotations, have emerged as a promising alternative. These SSL strategies are rapidly narrowing the performance gap with supervised learning methods. In this context, we introduce an SSL framework that dynamically adjusts the selection of negative samples during contrastive learning within the SimCLR process and incorporates masked autoencoders to reconstruct original images using latent representations and masked annotations. Notably, our approach outperforms common SSL methods in downstream classification tasks, validated through testing on the PathMNIST and our collected lung squamous carcinoma datasets.