Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Theory and GNNs to Unravel the Topographical Organization of Brain Lesions in Variants of Alzheimer's Disease Progression

Gabriel Jimenez, Leopold Hebert-Stevens, Benoit Delatour, Lev Stimmer, Daniel Racoceanu

TL;DR

This paper addresses the heterogeneity of Alzheimer's disease progression by leveraging graph representations of tau pathology extracted from high-resolution histopathology images. It combines a graph-based pipeline with explainable ML and multiple GNN architectures to distinguish between cAD and rpAD, revealing that rpAD forms denser topological networks and predominantly affects middle cortical layers, while cAD impacts outer layers. Key contributions include a tau-pathology graph construction via Delaunay triangulation with alpha-controlled erosion, two graph-scale representations (patient- and layer-level), and a hybrid analytical framework using RF with SHAP alongside GNN-derived embeddings and explainability tools (GNNExplainer, PGExplainer). Findings show plaque- and tangle-based graphs differ in discriminative power, with tangle graphs at the patient level and plaque embeddings at the layer level yielding strong classification performance, suggesting distinct neuropathological networks underpinning AD variants. While demonstrated on a small dataset, the work provides a methodological framework that integrates topology, graph learning, and XAI to advance understanding of AD progression and supports future expansion to larger cohorts and mechanistic investigations, potentially informing precision diagnostics and targeted interventions.

Abstract

In this study, we proposed and evaluated a graph-based framework to assess variations in Alzheimer's disease (AD) neuropathologies, focusing on classic (cAD) and rapid (rpAD) progression forms. Histopathological images are converted into tau-pathology-based (i.e., amyloid plaques and tau tangles) graphs, and derived metrics are used in a machine-learning classifier. This classifier incorporates SHAP value explainability to differentiate between cAD and rpAD. Furthermore, we tested graph neural networks (GNNs) to extract topological embeddings from the graphs and use them in classifying the progression forms of AD. The analysis demonstrated denser networks in rpAD and a distinctive impact on brain cortical layers: rpAD predominantly affects middle layers, whereas cAD influences both superficial and deep layers of the same cortical regions. These results suggest a unique neuropathological network organization for each AD variant.

Graph Theory and GNNs to Unravel the Topographical Organization of Brain Lesions in Variants of Alzheimer's Disease Progression

TL;DR

This paper addresses the heterogeneity of Alzheimer's disease progression by leveraging graph representations of tau pathology extracted from high-resolution histopathology images. It combines a graph-based pipeline with explainable ML and multiple GNN architectures to distinguish between cAD and rpAD, revealing that rpAD forms denser topological networks and predominantly affects middle cortical layers, while cAD impacts outer layers. Key contributions include a tau-pathology graph construction via Delaunay triangulation with alpha-controlled erosion, two graph-scale representations (patient- and layer-level), and a hybrid analytical framework using RF with SHAP alongside GNN-derived embeddings and explainability tools (GNNExplainer, PGExplainer). Findings show plaque- and tangle-based graphs differ in discriminative power, with tangle graphs at the patient level and plaque embeddings at the layer level yielding strong classification performance, suggesting distinct neuropathological networks underpinning AD variants. While demonstrated on a small dataset, the work provides a methodological framework that integrates topology, graph learning, and XAI to advance understanding of AD progression and supports future expansion to larger cohorts and mechanistic investigations, potentially informing precision diagnostics and targeted interventions.

Abstract

In this study, we proposed and evaluated a graph-based framework to assess variations in Alzheimer's disease (AD) neuropathologies, focusing on classic (cAD) and rapid (rpAD) progression forms. Histopathological images are converted into tau-pathology-based (i.e., amyloid plaques and tau tangles) graphs, and derived metrics are used in a machine-learning classifier. This classifier incorporates SHAP value explainability to differentiate between cAD and rpAD. Furthermore, we tested graph neural networks (GNNs) to extract topological embeddings from the graphs and use them in classifying the progression forms of AD. The analysis demonstrated denser networks in rpAD and a distinctive impact on brain cortical layers: rpAD predominantly affects middle layers, whereas cAD influences both superficial and deep layers of the same cortical regions. These results suggest a unique neuropathological network organization for each AD variant.
Paper Structure (10 sections, 3 figures, 1 table)

This paper contains 10 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Dataset and Graph Construction
  3. Methods
  4. Explainable Machine Learning
  5. Deep Learning Methods
  6. GNN Classifier:
  7. Clustering on node embeddings:
  8. GNN Explainer:
  9. Results and discussion
  10. Conclusion

Figures (3)

  • Figure 1: Example of a WSI and its annotations on the Region of Interest (ROI) on the left and the 6 layers delimitation on the right. Plaques are shown in a red mask, and tangles in a blue mask.
  • Figure 2: Mean node and edges importance for cAD and rpAD for GNNExplainer.
  • Figure 3: Comparison of clustering method for raw data points and GNN-based embeddings. The top graph presents clustered nodes based on their spatial coordinates. In contrast, the bottom graph leverages GNN embeddings to cluster nodes.