Rethinking Multiple Instance Learning for Whole Slide Image Classification: A Good Instance Classifier is All You Need

TL;DR

This work tackles weakly supervised whole slide image classification by addressing both instance- and bag-level predictions in a unified MIL framework. It introduces instance-level weakly supervised contrastive learning (IWSCL) to learn discriminative instance representations and prototype-based pseudo-label generation (PPLG) to produce high-quality instance labels, all within a joint training loop guided by true negative instances. Across synthetic and real histopathology datasets, INS achieves state-of-the-art performance on both instance and bag classifications and provides interpretable heatmaps that reveal clinically meaningful patterns. The approach promises improved localization and decision support in pathology under slide-level supervision, enabling more granular insights and potential discovery of novel patterns in medical images.

Abstract

Weakly supervised whole slide image classification is usually formulated as a multiple instance learning (MIL) problem, where each slide is treated as a bag, and the patches cut out of it are treated as instances. Existing methods either train an instance classifier through pseudo-labeling or aggregate instance features into a bag feature through attention mechanisms and then train a bag classifier, where the attention scores can be used for instance-level classification. However, the pseudo instance labels constructed by the former usually contain a lot of noise, and the attention scores constructed by the latter are not accurate enough, both of which affect their performance. In this paper, we propose an instance-level MIL framework based on contrastive learning and prototype learning to effectively accomplish both instance classification and bag classification tasks. To this end, we propose an instance-level weakly supervised contrastive learning algorithm for the first time under the MIL setting to effectively learn instance feature representation. We also propose an accurate pseudo label generation method through prototype learning. We then develop a joint training strategy for weakly supervised contrastive learning, prototype learning, and instance classifier training. Extensive experiments and visualizations on four datasets demonstrate the powerful performance of our method. Codes are available at https://github.com/miccaiif/INS.

Figures (7)

• Figure 1: Motivation of our method. (A). Existing instance-based MIL methods typically assign a bag's label to its instances as pseudo labels, resulting in a large number of noises in instance pseudo labels. (B). The loss function of bag-based MIL methods is defined at the bag level, which often only finds the most easily identifiable positive instances and ignores other more difficult ones. (C). In contrast to these methods, we propose an effective instance-based MIL framework based on contrastive learning and prototype learning and a joint training strategy. Our main goal is to develop an effective instance classifier using instance-level weakly supervised contrastive learning and pseudo-label generation. Contrastive learning aims to improve instance representations to better distinguish between negative and positive instances in the feature space. High-quality pseudo-labels are then generated for each instance based on representative feature vectors. Both feature learning and pseudo-label generation are guided by the instance classifier, enhancing its capabilities.
• Figure 2: Comparison of our instance-level weakly supervised contrastive learning (IWSCL)
• Figure 3: Workflow of the proposed INS framework, where "pos" is the abbreviation for positive, "neg" for negative, "Aug" for augmentation, "IWSCL" for the proposed instance-level weakly supervised contrastive learning, and "PPLG" for the proposed prototype-based pseudo label generation.
• Figure 4: Typical visualization results on the Camelyon 16 Dataset, where the yellow line represents the true tumor boundary annotated by doctors, and the pick boxes represent the positive instances predicted by INS as heatmaps.
• Figure 5: Heatmap visualization of INS on the lymph node metastasis task, where we identify the "Micropapillae" pathological pattern that indicates positive metastasis from HE slides.