RNN-Based Multiple Instance Learning for the Classification of Histopathology Whole Slide Images

The digitization of histopathology Whole Slide Images (WSIs) enables deep learning to be applied in pathology research and medical computer-aided diagnosis. Due to the high-resolution of WSIs, multiple-instance learning (MIL) has become a powerful tool for addressing weakly supervised WSI classification tasks. However, most of the existing MIL approaches ignore the dependency of instances, limiting the adequate extraction of global WSI representation. Further, MIL approaches are often based on attention mechanisms, usually with huge time and space complexity. To tackle these problems, we propose an RWKV-based MIL, adapted from an RNN-like NLP model. Its instance mixing mechanism can model the long-range dependency of instances. And it can work parallelly with lower time and space complexity. As a result, our method can efficiently learn global representations from high-resolution WSIs. Compared with the original RWKV, we propose various new designs to achieve the best performance in WSI classification. Experiments and ablation studies are conducted on two public WSI datasets with four baselines to demonstrate the better performance of our RWKV-based MIL. The AUC results of our proposed method on CMELYON16 and TCGA-BRCA are 0.7964 and 0.8971, gaining an improvement of 7.7% and 0.8% over existing methods, respectively.