Document Type

Article

Publication Date

12-2025

Publisher

Institute of Electrical and Electronics Engineers (IEEE)

Source Publication

IEEE Transactions on Artificial Intelligence

Source ISSN

2691-4581

Original Item ID

DOI: 10.1109/TAI.2025.3646956

Abstract

High-resolution digital scans of pathology slides, known as whole slide images (WSIs), have detailed spatial and contextual information for diagnosing cancer. However, the classification performance of WSIs by deep learning models is typically compromised by data with a different distribution, known as out-of-distribution (OOD), resulting in unreliable predictions. Therefore, having a reliable predictive uncertainty estimation is crucial for clinical adoption. This article comprehensively studies graph-based uncertainty estimation for WSI classification using two cutting-edge graph neural network (GNN) architectures: 1) graph attention networks (GAT); and 2) GraphSAGE. In this work, we introduce the first unified multihead GNN framework that leverages GraphSAGE backbones for predictive uncertainty estimation. We propose and evaluate novel multihead GNN frameworks (MH-GAT and MH-GraphSAGE) that use multiple output branches to improve model robustness and quantify predictive uncertainty via head-wise divergence. Our method is evaluated on a variety of datasets, including binary and multiclass breast cancer classification, in both in-distribution (ID) and OOD settings. Experiments demonstrate that multihead GNNs consistently outperform vanilla single-head, dropout-based, and ensemble-based approaches in both classification accuracy and uncertainty estimation. We also introduce WSI-level uncertainty visualizations—such as heatmaps and ranking plots—which can help identify ambiguous or error-prone cases for clinical review and interpretability. Our findings confirm the use of multihead GNNs as an accurate, uncertainty-aware approach for digital pathology applications.

Comments

Accepted version. IEEE Transactions on Artificial Intelligence, (December 2025): 1- 11. DOI. © 2025 IEEE. Used with permission.

Download
Yahyasoltani_17676acc.docx (943 kB)
ADA Accessible Version

Share

COinS