NcIEMIL: Rethinking Decoupled Multiple Instance Learning Framework for Histopathological Slide Classification

Sun Qiehe, Doukou Jiang, Jiawen Li, Renao Yan, Yonghong He, Tian Guan, Zhiqiang Cheng

Keywords: Multiple instance learning, Histopathological slide classification, Instance cleansing, Hybrid attention-based aggregation

Abstract: On account of superiority in annotation efficiency, multiple instance learning (MIL) has proved to be a promising framework for the whole slide image (WSI) classification in pathological diagnosis. However, current methods employ fully- or semi-decoupled frameworks to address the trade-off between billions of pixels and limited computational resources. This exacerbates the information bottleneck, leading to instance representations in a high-rank space that contains semantic redundancy compared to the potential low-rank category space of instances. Additionally, most negative instances are also independent of the positive properties of the bag. To address this, we introduce a weakly annotation-supervised filtering network, aiming to restore the low-rank nature of the slide-level representations. We then design a parallel aggregation structure that utilizes spatial attention mechanisms to model inter-correlation between instances and simultaneously assigns corresponding weights to channel dimensions to alleviate the redundant information introduced by feature extraction. Extensive experiments on the private gastrointestinal chemotaxis dataset and CAMELYON16 breast dataset show that our proposed framework is capable of handling both binary and multivariate classification problems and outperforms state-of-the-art MIL-based methods. The code is available at: https://github.com/polyethylene16/NcIEMIL.

Submission Number: 99