Multi-Instance Learning for Whole-Slide Image Classification Using Higher-Order Moments

Xia Zhixiang, wuji, Xiaofan Wu, Guosheng Yin, Bin Liu

Keywords: Multi-Instance Learning, Whole-Slide Image, Attention, High-Order Moments

Abstract: Whole-slide images (WSIs) contain abundant pathological information. However, the extremely high resolution and substantial redundant information in WSIs pose significant challenges for both manual analysis and artificial intelligence processing. Multi-instance learning (MIL) is currently the predominant approach, which typically focuses on aggregating low-dimensional feature representations of all patches into a single vector. If the vectors of patches are regarded as random variables, this aggregation process is essentially equivalent to estimating the first-order moment of these random vectors. However, the first-order moment alone cannot fully capture the information of the entire slide, necessitating the computation of second-order moments. Specifically, we first employ attention-based multiple instance learning (ABMIL) to calculate the attention-weighted average of patches as an estimate of the first-order moment. Concurrently, we compute the covariance matrix of the patch representation vectors across the entire slide. By aggregating the information from both the first- and second-order moments, we can greatly enhance the classification accuracy of WSIs. To improve computational efficiency, we employ DBSCAN clustering that adaptively forms large clusters for abundant normal tissues and small clusters for rare pathological regions, enabling variable-resolution processing that preserves diagnostic information while reducing computational cost. Experimental results on multiple real-world datasets demonstrate that our model significantly improves the state-of-the-art performance.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 9121