Multiple instance learning with hierarchical discrimination and smoothing attention for histopathological diagnosis

The microscopic structure of human tissue can be observed by pathological slides, which provides a strong basis for cancer diagnosis. However, the serious lack of experienced pathologists and the complexity of the diagnostic process have facilitated the development of computer-aided pathological image analysis. Pathological slides generally have high resolution, and multiple instance learning (MIL) has been widely used in histopathological whole slide image (WSI) analysis, where each WSI has a large number of unlabelled patches and only a WSI-level label is given. The bag-based MIL methods often learn the decision boundary at the bag level, and thus hard to learn the discriminative features at the instance level. Furthermore, the difficulty of identification varies between positive instances in a bag, and the existing attention-based aggregation methods always assign higher attention scores for the easy-to-identify positive instances, but assign lower attention scores for the difficult-to-identify positive instances and thus cannot learn these difficult instances sufficiently. In this paper, we propose a novel MIL method with hierarchical discrimination learning and a smoothing attention strategy for cancer subtype diagnosis. Particularly, to learn hierarchical discriminative features, the proposed MIL method simultaneously trains a bag classifier and multiple instance classifiers, where the multi-way attention scores of each instance for different categories are used to guide the selection of training samples for the instance classifimer. The smoothing strategy is designed to trade off the attention weights between the easily and hardly identifiable positive instances. We conducted experiments on histopathological diagnosis datasets and achieved state-of-the-art performance. Codes are available at https://github.com/bravePinocchio/HDSA-MIL.