Norma: A Noise Robust Memory-Augmented Framework for Whole Slide Image Classification
Yu Bai, Bo Zhang*, Zheng Zhang, Shuo Yan, Zibo Ma, Wu Liu, Xiuzhuang Zhou, Xiangyang Gong, Wendong Wang
;
Abstract
"In recent years, the Whole Slide Image (WSI) classification task has achieved great advancement due to the success of Multiple Instance Learning (MIL). However, the MIL-based studies usually consider instances within each bag as unordered, potentially resulting in the missing of local and global contextual information. To overcome this limitation, we propose a Noise Robust Memory-Augmented (Norma) framework for addressing the WSI classification task using a sequential approach. Norma serializes a WSI into a long sequence and adopts the Vision Transformer (ViT) to encode the local and global context information of the WSIs. Instead of processing long sequences at once, Norma splits the long sequence into multiple segments and sequentially trains these segments, with each segment being cached for future reuse. In addition, considering that segment-level labels are inherited from slide-level labels, which may introduce noise during training, Norma further introduces a cyclic method to reduce label noise. We achieve state-of-the-art performance on the CAMELYON-16, TCGA-BRAC and TCGA-LUNG datasets compared to recent studies. The code is available at https://github.com/weiaicunzai/Norma."
Related Material
[pdf]
[supplementary material]
[DOI]