Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ProtoDiv: Prototype-guided Division of Consistent Pseudo-bags for Whole-slide Image Classification

Rui Yang, Pei Liu, Luping Ji

TL;DR

ProtoDiv tackles the data scarcity and weak labeling challenge in whole-slide image (WSI) classification by augmenting training data with pseudo-bags that are guided by bag prototypes. It introduces mean-based and attention-based prototypes to produce a cosine-distance–driven partitioning of patch features, yielding phenotype-diverse yet representative pseudo-bags that preserve the parent bag distribution. The approach is designed as a plug-and-play augmentation for existing MIL methods and demonstrates consistent AUC gains across seven baselines on TCGA-Lung and TCGA-BRCA, with efficient division times and intuitive visualizations of prototype separation. Overall, ProtoDiv enhances generalization in WSI MIL by maintaining phenotype consistency while safely increasing data size through adaptive pseudo-bag generation.

Abstract

Due to the limitations of inadequate Whole-Slide Image (WSI) samples with weak labels, pseudo-bag-based multiple instance learning (MIL) appears as a vibrant prospect in WSI classification. However, the pseudo-bag dividing scheme, often crucial for classification performance, is still an open topic worth exploring. Therefore, this paper proposes a novel scheme, ProtoDiv, using a bag prototype to guide the division of WSI pseudo-bags. Rather than designing complex network architecture, this scheme takes a plugin-and-play approach to safely augment WSI data for effective training while preserving sample consistency. Furthermore, we specially devise an attention-based prototype that could be optimized dynamically in training to adapt to a classification task. We apply our ProtoDiv scheme on seven baseline models, and then carry out a group of comparison experiments on two public WSI datasets. Experiments confirm our ProtoDiv could usually bring obvious performance improvements to WSI classification.

ProtoDiv: Prototype-guided Division of Consistent Pseudo-bags for Whole-slide Image Classification

TL;DR

ProtoDiv tackles the data scarcity and weak labeling challenge in whole-slide image (WSI) classification by augmenting training data with pseudo-bags that are guided by bag prototypes. It introduces mean-based and attention-based prototypes to produce a cosine-distance–driven partitioning of patch features, yielding phenotype-diverse yet representative pseudo-bags that preserve the parent bag distribution. The approach is designed as a plug-and-play augmentation for existing MIL methods and demonstrates consistent AUC gains across seven baselines on TCGA-Lung and TCGA-BRCA, with efficient division times and intuitive visualizations of prototype separation. Overall, ProtoDiv enhances generalization in WSI MIL by maintaining phenotype consistency while safely increasing data size through adaptive pseudo-bag generation.

Abstract

Due to the limitations of inadequate Whole-Slide Image (WSI) samples with weak labels, pseudo-bag-based multiple instance learning (MIL) appears as a vibrant prospect in WSI classification. However, the pseudo-bag dividing scheme, often crucial for classification performance, is still an open topic worth exploring. Therefore, this paper proposes a novel scheme, ProtoDiv, using a bag prototype to guide the division of WSI pseudo-bags. Rather than designing complex network architecture, this scheme takes a plugin-and-play approach to safely augment WSI data for effective training while preserving sample consistency. Furthermore, we specially devise an attention-based prototype that could be optimized dynamically in training to adapt to a classification task. We apply our ProtoDiv scheme on seven baseline models, and then carry out a group of comparison experiments on two public WSI datasets. Experiments confirm our ProtoDiv could usually bring obvious performance improvements to WSI classification.
Paper Structure (16 sections, 5 equations, 5 figures, 3 tables)

This paper contains 16 sections, 5 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. Problem Formulation
  4. Traditional MIL Settings.
  5. Integrating MIL with ProtoDiv.
  6. Our ProtoDiv Scheme
  7. Prototype Calculation.
  8. Bag Dividing via Prototype.
  9. Experiment and Result
  10. Experimental Settings.
  11. Implementation Details.
  12. Classification Performance.
  13. Ablation Study.
  14. Result Visualization.
  15. Discussion
  16. ...and 1 more sections

Figures (5)

  • Figure 1: An overview of ProtoDiv. One WSI is used as an example to explain how our ProtoDiv works in MIL framework: extract patch features, calculate bag prototype, divide bag into pseudo-bags for training, and aggregate to get a final prediction.
  • Figure 2: Visualization of the attention prototypes in TCGA-BRCA (left) and TCGA-Lung (right) test set by using t-SNE vandermaaten08a. ABMIL is adopted as the baseline model.
  • Figure 3: Visualization of the phenotype and pseudo-bag generated by our ProtoDiv. (a) is the original WSI. (b) shows the cluster prototypes via ProtoDiv. (c) gives the pseudo-bags generated via ProtoDiv. Each color indicates one specific phenotype or pseudo-bag in (b) and (c). More results are shown in our supplementary material.
  • Figure 4: All phenotypes and their respective patches, as a supplement to the Figure 3 in our main context. Patches are clustered into different tissue phenotypes for subsequent pseudo-bag generation.
  • Figure 5: Visualization of the phenotype and pseudo-bag generated by our ProtoDiv. (a) is the original WSI. (b) shows the cluster prototypes via ProtoDiv. (c) gives the pseudo-bags generated via ProtoDiv. Each color indicates one specific phenotype or pseudo-bag in (b) and (c), respectively. The patches of three selected phenotypes are also shown at the bottom of each row.