Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Integrating multiscale topology in digital pathology with pyramidal graph convolutional networks

Victor Ibañez, Przemyslaw Szostak, Quincy Wong, Konstanty Korski, Samaneh Abbasi-Sureshjani, Alvaro Gomariz

TL;DR

The paper tackles the challenge of deriving rich spatial context from gigapixel WSIs by overcoming oversmoothing in single-magnification GCNs. It introduces MS-GCN, a multiscale graph neural network that constructs a pyramid of graphs across magnifications $m \in \{1,\dots,M\}$ and fuses information via $L = M-1$ stacked $\mathcal{F}_{\mathrm{GEN}}$ layers with a global attention pooling, producing a multiscale embedding $\mathbf{z}$. By decomposing $\mathbf{z}$ into magnification-specific components, the method provides influence scores that reveal which scales drive predictions, enhancing interpretability. Empirical results on four diverse datasets show MS-GCN consistently outperforms single-magnification Patch-GCN, with notable gains in breast cancer grading, while attention analyses confirm scale-dependent feature contributions and pathologist-like interpretability. The work advances computational pathology by delivering both improved accuracy and a principled avenue for cross-scale interpretability, albeit with increased computational demands.

Abstract

Graph convolutional networks (GCNs) have emerged as a powerful alternative to multiple instance learning with convolutional neural networks in digital pathology, offering superior handling of structural information across various spatial ranges - a crucial aspect of learning from gigapixel H&E-stained whole slide images (WSI). However, graph message-passing algorithms often suffer from oversmoothing when aggregating a large neighborhood. Hence, effective modeling of multi-range interactions relies on the careful construction of the graph. Our proposed multi-scale GCN (MS-GCN) tackles this issue by leveraging information across multiple magnification levels in WSIs. MS-GCN enables the simultaneous modeling of long-range structural dependencies at lower magnifications and high-resolution cellular details at higher magnifications, akin to analysis pipelines usually conducted by pathologists. The architecture's unique configuration allows for the concurrent modeling of structural patterns at lower magnifications and detailed cellular features at higher ones, while also quantifying the contribution of each magnification level to the prediction. Through testing on different datasets, MS-GCN demonstrates superior performance over existing single-magnification GCN methods. The enhancement in performance and interpretability afforded by our method holds promise for advancing computational pathology models, especially in tasks requiring extensive spatial context.

Integrating multiscale topology in digital pathology with pyramidal graph convolutional networks

TL;DR

The paper tackles the challenge of deriving rich spatial context from gigapixel WSIs by overcoming oversmoothing in single-magnification GCNs. It introduces MS-GCN, a multiscale graph neural network that constructs a pyramid of graphs across magnifications and fuses information via stacked layers with a global attention pooling, producing a multiscale embedding . By decomposing into magnification-specific components, the method provides influence scores that reveal which scales drive predictions, enhancing interpretability. Empirical results on four diverse datasets show MS-GCN consistently outperforms single-magnification Patch-GCN, with notable gains in breast cancer grading, while attention analyses confirm scale-dependent feature contributions and pathologist-like interpretability. The work advances computational pathology by delivering both improved accuracy and a principled avenue for cross-scale interpretability, albeit with increased computational demands.

Abstract

Graph convolutional networks (GCNs) have emerged as a powerful alternative to multiple instance learning with convolutional neural networks in digital pathology, offering superior handling of structural information across various spatial ranges - a crucial aspect of learning from gigapixel H&E-stained whole slide images (WSI). However, graph message-passing algorithms often suffer from oversmoothing when aggregating a large neighborhood. Hence, effective modeling of multi-range interactions relies on the careful construction of the graph. Our proposed multi-scale GCN (MS-GCN) tackles this issue by leveraging information across multiple magnification levels in WSIs. MS-GCN enables the simultaneous modeling of long-range structural dependencies at lower magnifications and high-resolution cellular details at higher magnifications, akin to analysis pipelines usually conducted by pathologists. The architecture's unique configuration allows for the concurrent modeling of structural patterns at lower magnifications and detailed cellular features at higher ones, while also quantifying the contribution of each magnification level to the prediction. Through testing on different datasets, MS-GCN demonstrates superior performance over existing single-magnification GCN methods. The enhancement in performance and interpretability afforded by our method holds promise for advancing computational pathology models, especially in tasks requiring extensive spatial context.
Paper Structure (13 sections, 3 equations, 2 figures, 2 tables)

This paper contains 13 sections, 3 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Multi-scale graph construction
  4. MS-GCN architecture
  5. Influence scores for multiscale interpretability
  6. Results and Discussion
  7. Dataset and implementation details
  8. Evaluating the performance of MS-GCN
  9. Discriminating tissue level information
  10. Interpretability of multi-scale graph features
  11. Conclusion
  12. Compliance with Ethical Standards
  13. Acknowledgments

Figures (2)

  • Figure 1: Illustration of proposed MS-GCN method. (top) Higher magnification vertices (color coded) are connected to lower magnification ones recursively. Only the lowest resolution vertices are connected spatially among them. (middle) Message passage (magenta) algorithm in a graph following the color codes from top. (bottom) Magnification-specific attention heatmaps.
  • Figure 2: Attention heatmaps for the different datasets (rows). The columns show the raw WSI as well as the attention heatmaps at different magnifications as overlays on the WSI.