DGMIL: Distribution Guided Multiple Instance Learning for Whole Slide Image Classification

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 pseudolabeling 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 instancelevel 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 instancelevel 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 will be available.

Contrastive Multiple Instance Learning: An Unsupervised Framework for Learning Slide-Level Representations of Whole Slide Histopathology Images without Labels

“…This was notable at the time, as previous methods had relied on tediously annotated tissue ROIs. Several studies have since followed suit [ 4 , 17 , 20 , 25 , 67 , 68 , 69 , 70 ].…”

“…Courtiol et al [ 65 ] were the first to apply MIL to subtype cancer using WSIs—specifically, subtyping NSCLC into LUAD and LUSC using WSIs available from TCGA [ 71 ]—and drew similar conclusions and implications from their work with Camelyon16. The standard set by Courtiol et al has endured through subsequent studies [ 17 , 25 , 68 , 69 , 70 , 72 , 73 , 74 ]. Wang et al added an addition subtype for small-cell lung cancer (SCLC) [ 43 ] using an in-house dataset.…”

Background detection affects downstream classification of Camelyon16 whole slide images

“…We posit that such variation in background detection methods is one of the causes of the inability to reproduce specific methods' results on specific datasets as reported in Table 2. [19] 0.8938 0.8832 0.8939 DGMIL [20] 0.7544 0.6612 0.6913 DTFD-MIL [21] 0.906 0.878 0.899 0.854 0.875 A2M [2] 0.858 0.588 0.740 TransMIL [4] 0.9309 0.8679 0.8179 0.876 0.888 0.8569 0.6389 DS-MIL [14] 0.8944 0.8653 0.8064 0.8641 0.762 CLAM [12] 0.936 Hashimoto [22] 0.8895 0.9165 0.84 0.8064 MIL-RNN [15] 0.899 Courtiol [23] 0.655 0.53…”

The effects of sparsity induction methods on attention-based multiple instance learning applied to Camelyon16

“…There are several weakly-supervised methods for classification of WSIs that may similarly benefit from these sparsityinduction methods. MIL-RNN [9], CLAM [10], DS-MIL [11], TransMIL [12], Zhang et al [3], and DTFD-MIL [13] reported AUCs of 0.899, 0.936, 0.8944, 0.9309, 0.9377, and 0.946, respectively, although several studies have failed to reproduce the former four results [11,12,[14][15][16][17][18]. With the exception of the study by Zhang et al, none of these studies employed the reported sparsity induction methods.…”