RCNet: A Redundant Compression Network Using Information Bottleneck for Pathology Whole Slide Image Classification
Hongxuan Yu, Jiayi Wu, Jichen Xu, Shuhao Wang, Wei Wang, Siyi Chai, Jingmin Xin · 2024
The analysis of Whole Slide Images (WSIs) is vital for tumor diagnosis, and numerous deep learning methods have been extensively researched. However, the large size and multi-scale features of WSIs introduce substantial redundant information, leading to significant performance challenges. Existing methods have struggled to address this redundancy effectively. To reduce redundant information, we propose a novel Redundant Compression Network (RCNet) for WSI classification, incorporating an effective Multi-Scale Fusion Module (MSFM) and an Information Bottleneck Compression Module (IBCM). Specifically, the MSFM filters the redundancy in multi-scale features by emphasizing critical scales, while the IBCM eliminates redundant instances through compression using the information bottleneck technique. We evaluate our method on two datasets, the public DigestPath2019 dataset and a private lung pathology dataset. On the public dataset, our approach achieves improvements of at least 2.4% in F1-score, 2.3% in accuracy, 3.8% in Recall and 1.0% in AUC compared to state-of-the-art methods. Additionally, on the private dataset, our approach achieves improvements of at least 1.6% in F1-score, 1.2% in accuracy and 0.8% in AUC. These results demonstrate the effectiveness of our method in improving WSI classification by reducing redundancy and focusing on the most essential information.