Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting ulcer in H&E images of inflammatory bowel disease using domain-knowledge-driven graph neural network

Ruiwen Ding, Lin Li, Rajath Soans, Tosha Shah, Radha Krishnan, Marc Alexander Sze, Sasha Lukyanov, Yash Deshpande, Antong Chen

TL;DR

The study tackles ulcer detection in IBD histology WSIs by introducing DomainGCN, a domain-knowledge-driven GNN that constructs WSI graphs from patches, fusing high-level UNI features with positional encoding and weighting nodes by ulcer-relevant tissue content. By injecting domain knowledge—epithelium, lymphocytes, and debris probabilities—into the message-passing process, DomainGCN enhances WSI-level ulcer prediction and yields more localized, plausible attention to ulcer regions. Empirical results show graph-based methods outperform MIL approaches, with DomainGCN achieving the highest mean AUC (0.800), F1 (0.794), and ACC (0.793), and domain weighting further improving consistency. This approach enables spatially aware ulcer characterization and sets the stage for integrated immune-cell analysis, potentially generalizing to other GI tract locations.

Abstract

Inflammatory bowel disease (IBD) involves chronic inflammation of the digestive tract, with treatment options often burdened by adverse effects. Identifying biomarkers for personalized treatment is crucial. While immune cells play a key role in IBD, accurately identifying ulcer regions in whole slide images (WSIs) is essential for characterizing these cells and exploring potential therapeutics. Multiple instance learning (MIL) approaches have advanced WSI analysis but they lack spatial context awareness. In this work, we propose a weakly-supervised model called DomainGCN that employs a graph convolution neural network (GCN) and incorporates domain-specific knowledge of ulcer features, specifically, the presence of epithelium, lymphocytes, and debris for WSI-level ulcer prediction in IBD. We demonstrate that DomainGCN outperforms various state-of-the-art (SOTA) MIL methods and show the added value of domain knowledge.

Predicting ulcer in H&E images of inflammatory bowel disease using domain-knowledge-driven graph neural network

TL;DR

The study tackles ulcer detection in IBD histology WSIs by introducing DomainGCN, a domain-knowledge-driven GNN that constructs WSI graphs from patches, fusing high-level UNI features with positional encoding and weighting nodes by ulcer-relevant tissue content. By injecting domain knowledge—epithelium, lymphocytes, and debris probabilities—into the message-passing process, DomainGCN enhances WSI-level ulcer prediction and yields more localized, plausible attention to ulcer regions. Empirical results show graph-based methods outperform MIL approaches, with DomainGCN achieving the highest mean AUC (0.800), F1 (0.794), and ACC (0.793), and domain weighting further improving consistency. This approach enables spatially aware ulcer characterization and sets the stage for integrated immune-cell analysis, potentially generalizing to other GI tract locations.

Abstract

Inflammatory bowel disease (IBD) involves chronic inflammation of the digestive tract, with treatment options often burdened by adverse effects. Identifying biomarkers for personalized treatment is crucial. While immune cells play a key role in IBD, accurately identifying ulcer regions in whole slide images (WSIs) is essential for characterizing these cells and exploring potential therapeutics. Multiple instance learning (MIL) approaches have advanced WSI analysis but they lack spatial context awareness. In this work, we propose a weakly-supervised model called DomainGCN that employs a graph convolution neural network (GCN) and incorporates domain-specific knowledge of ulcer features, specifically, the presence of epithelium, lymphocytes, and debris for WSI-level ulcer prediction in IBD. We demonstrate that DomainGCN outperforms various state-of-the-art (SOTA) MIL methods and show the added value of domain knowledge.

Paper Structure

This paper contains 9 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset
  4. Graph construction and model architecture
  5. Incorporation of domain knowledge (ulcer weight)
  6. Baseline models
  7. Model training and evaluation
  8. Results and Discussion
  9. Compliance with ethical standards

Figures (3)

  • Figure 1: The pipeline of DomainGCN including WSI patching, node feature extraction, 8-nearest-neighbor graph construction, and graph modeling.
  • Figure 2: Boxplots showing the validation AUC of all models compared in this study.
  • Figure 3: Visualization of domain knowledge and attention maps. (a) represents boxplots showing the distribution of ulcer weights for non-ulcer and ulcer cases. (b) visualizes the patches that have ulcer weight 3 or 4. (c)(d)(e) respectively shows the patches with the top 25% attention scores for DomainGCN, AMIl, and CLAM. (f) and (g) show the two ulcer regions of this case.