Second-Order Self-Supervised Learning for Breast Cancer Classification

Haitao Yao, Zhenwei Wang, Mingli Zhang, Wen Zhu, Lizhi Zhang, Lijun He, Jianxin Zhang

Abstract: Multiple instance learning (MIL), which adopts instance-level feature encoders in a weakly supervised manner, efficiently classifies cancers on whole slide images (WSis). Nevertheless, the lack of fine-grained annotation samples still limits the performance of existing MIL models. Therefore, we propose a second-order self-supervised learning method integrated with the MIL structure, called SoS2MIL, for breast cancer classification. Specifically, SoS2MIL explores intrinsic structural information within WSI instances using self-supervised learning that does not require label information to identify discriminative instances. Then, it improves the instance-level classifier concentration by introducing second-order pooling in a visual perception module. Besides, a grouping strategy is proposed to reduce the computational cost of second-order calculations and effectively approximate feature distributions by combining first-order features. Experimental results on public and private datasets verify its effectiveness and competitiveness with the state-of-the-art MIL works.