Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WSD-MIL: Window Scale Decay Multiple Instance Learning for Whole Slide Image Classification

Le Feng, Li Xiao

TL;DR

WSD-MIL tackles scalable, weakly supervised WSI classification by introducing window-scale decay attention (WSDA) and a squeeze-and-excitation region gate (SERG). WSDA employs cluster-based sampling and a Nyström-augmented multi-scale window transformer to capture local correlations across tumor regions of varying sizes while reducing $O(n^2)$ attention costs. SERG provides global inter-region weighting to emphasize discriminative regions, yielding a global-to-local-to-global feature refinement. Across CAMELYON-16 and TCGA-BRCA, WSD-MIL achieves state-of-the-art results with substantial memory savings (about $62\%$) and strong synergy with foundation-model embeddings, highlighting its practical impact for scalable computational pathology.

Abstract

In recent years, the integration of pre-trained foundational models with multiple instance learning (MIL) has improved diagnostic accuracy in computational pathology. However, existing MIL methods focus on optimizing feature extractors and aggregation strategies while overlooking the complex semantic relationships among instances within whole slide image (WSI). Although Transformer-based MIL approaches aiming to model instance dependencies, the quadratic computational complexity limits their scalability to large-scale WSIs. Moreover, due to the pronounced variations in tumor region scales across different WSIs, existing Transformer-based methods employing fixed-scale attention mechanisms face significant challenges in precisely capturing local instance correlations and fail to account for the distance-based decay effect of patch relevance. To address these challenges, we propose window scale decay MIL (WSD-MIL), designed to enhance the capacity to model tumor regions of varying scales while improving computational efficiency. WSD-MIL comprises: 1) a window scale decay based attention module, which employs a cluster-based sampling strategy to reduce computational costs while progressively decaying attention window-scale to capture local instance relationships at varying scales; and 2) a squeeze-and-excitation based region gate module, which dynamically adjusts window weights to enhance global information modeling. Experimental results demonstrate that WSD-MIL achieves state-of-the-art performance on the CAMELYON16 and TCGA-BRCA datasets while reducing 62% of the computational memory. The code will be publicly available.

WSD-MIL: Window Scale Decay Multiple Instance Learning for Whole Slide Image Classification

TL;DR

WSD-MIL tackles scalable, weakly supervised WSI classification by introducing window-scale decay attention (WSDA) and a squeeze-and-excitation region gate (SERG). WSDA employs cluster-based sampling and a Nyström-augmented multi-scale window transformer to capture local correlations across tumor regions of varying sizes while reducing attention costs. SERG provides global inter-region weighting to emphasize discriminative regions, yielding a global-to-local-to-global feature refinement. Across CAMELYON-16 and TCGA-BRCA, WSD-MIL achieves state-of-the-art results with substantial memory savings (about ) and strong synergy with foundation-model embeddings, highlighting its practical impact for scalable computational pathology.

Abstract

In recent years, the integration of pre-trained foundational models with multiple instance learning (MIL) has improved diagnostic accuracy in computational pathology. However, existing MIL methods focus on optimizing feature extractors and aggregation strategies while overlooking the complex semantic relationships among instances within whole slide image (WSI). Although Transformer-based MIL approaches aiming to model instance dependencies, the quadratic computational complexity limits their scalability to large-scale WSIs. Moreover, due to the pronounced variations in tumor region scales across different WSIs, existing Transformer-based methods employing fixed-scale attention mechanisms face significant challenges in precisely capturing local instance correlations and fail to account for the distance-based decay effect of patch relevance. To address these challenges, we propose window scale decay MIL (WSD-MIL), designed to enhance the capacity to model tumor regions of varying scales while improving computational efficiency. WSD-MIL comprises: 1) a window scale decay based attention module, which employs a cluster-based sampling strategy to reduce computational costs while progressively decaying attention window-scale to capture local instance relationships at varying scales; and 2) a squeeze-and-excitation based region gate module, which dynamically adjusts window weights to enhance global information modeling. Experimental results demonstrate that WSD-MIL achieves state-of-the-art performance on the CAMELYON16 and TCGA-BRCA datasets while reducing 62% of the computational memory. The code will be publicly available.
Paper Structure (15 sections, 14 equations, 4 figures, 3 tables)

This paper contains 15 sections, 14 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Problem formulation
  4. Overall architecture of window scale decay multiple instance learning framework (WSD-MIL)
  5. Window scale decay based attention module (WSDA)
  6. Feature Clustering-based Sampling Strategy
  7. Decaying Window Transformer
  8. Squeeze-and-excitation based region gate module (SERG).
  9. Experiments and Results
  10. Datasets
  11. Experiment setup and evaluation metrics
  12. Sample results
  13. Comparison with State-of-the-art methods
  14. Ablation study
  15. Conclusion

Figures (4)

  • Figure 1: Top: The previous methods employed fixed-scale attention, which struggled to accurately capture key features in WSIs with significant scale variations in tumor regions. Bottom: The proposed method introduces a scale-decaying attention, enabling the model to flexibly adapt to tumor regions of different scales and thus precisely extract key features.
  • Figure 2: Overview of the proposed WSD-MIL.
  • Figure 3: The structures of the decaying window transformer and the squeeze-excitation region gate.
  • Figure 4: The image preprocessing steps for WSIs. (a) Original WSI samples; (b)The foreground region of the WSI obtained via threshold-based segmentation; (c) The WSI after background removal; (d) Multiple 256 × 256-pixel patches obtained by cropping the foreground of the WSI.