Cross-Patient Pseudo Bags Generation and Curriculum Contrastive Learning for Imbalanced Multiclassification of Whole Slide Image

TL;DR

This work proposes learning fine-grained information by generating sub-bags with feature distributions similar to the original WSIs, using a pseudo-bag generation algorithm to further leverage the abundant and redundant information in WSIs, allowing efficient training in unbalanced-sample multi-classification tasks.

Abstract

Pathology computing has dramatically improved pathologists' workflow and diagnostic decision-making processes. Although computer-aided diagnostic systems have shown considerable value in whole slide image (WSI) analysis, the problem of multi-classification under sample imbalance remains an intractable challenge. To address this, we propose learning fine-grained information by generating sub-bags with feature distributions similar to the original WSIs. Additionally, we utilize a pseudo-bag generation algorithm to further leverage the abundant and redundant information in WSIs, allowing efficient training in unbalanced-sample multi-classification tasks. Furthermore, we introduce an affinity-based sample selection and curriculum contrastive learning strategy to enhance the stability of model representation learning. Unlike previous approaches, our framework transitions from learning bag-level representations to understanding and exploiting the feature distribution of multi-instance bags. Our method demonstrates significant performance improvements on three datasets, including tumor classification and lymph node metastasis. On average, it achieves a 4.39-point improvement in F1 score compared to the second-best method across the three tasks, underscoring its superior performance.

Figures (4)

• Figure 1: The motivation of our method. Left: Conventional methods aggregate all instances coarsely without considering further efficient utilization of all instances. Right: We propose a cross-patient instance utilization method based on feature distribution and a curriculum contrastive method to improve the expressive ability of the model and alleviate the issue of multi-class imbalance.
• Figure 2: Framework Overview: We process pathological images into multi-instance bags and generate sub-bags with similar feature distributions. These cross-patient sub-bags are combined to create pseudo-bags, leveraging the redundant information in WSIs. An affinity-based sample selection method is employed to achieve curriculum contrastive learning. These techniques collectively improve the model's performance in addressing multi-class imbalance in pathological tasks.
• Figure 3: Tumor localization in the whole slide images.
• Figure 4: Visualization of feature distribution on the TCGA Kidney using t-SNE.