BS2CL: Balanced Self-supervised Contrastive Learning for Thyroid Cytology Whole Slide Image Multi-classification

Thyroid cytology whole slide images (WSIs) hold vital information essential for precise diagnosis. Given the huge size of WSIs, multiple instance learning (MIL) is an effective solution for the WSI classification task when only slide-level labels are accessible. The embedding-based MIL uses a feature extractor pretrained with the self-supervised contrastive learning framework to eliminate the dependence on patch-level labels. However, the distribution of class in thyroid patches is unbalanced, and most existing self-supervised contrastive learning methods take little account of the data imbalance, which makes the features not discriminative enough. To address this problem, we propose a novel balanced self-supervised contrastive learning (BS2CL) framework for pretraining. It first clusters the patches to preserve the class structure of the patches and then assigns the clustering centers to a set of pre-computed uniformly distributed optimal locations. This constraint creates a more uniform distribution of different classes in the feature space which leads to clearer class boundaries between different classes, an unbiased feature space, and more discriminative features. Furthermore, a bag-level data augmentation strategy is introduced to increase bag quantities and improve classification performance. Extensive experiments show that the proposed method outperforms other latest methods on the thyroid cytology WSI dataset.