Atefeh Azin Kousha
Abstract: Graph Convolutional Networks (GCNs) perform best on homophilic graphs, where connected nodes share similar features. However, histopathology images present a challenge due to heterogeneous tissue patterns and diverse morphological structures. In such cases, ensuring homophily in the graph model is crucial, since in non-homophilic graphs message passing can lead to feature over-smoothing, where features from dissimilar tissue regions become mixed and reduce the classifier’s ability to distinguish distinct patterns. To address this, we introduce Boosted Adaptive Radius Graph (BARG), a novel graph modelling strategy tailored for Hematoxylin and Eosin (H&E)-stained histology images. BARG builds upon a Fixed Radius Graph (FRG) dataset in which a global distance threshold for nodes’ connectivity is optimized. Through statistical analysis of nodes’ Local Density Features (LDFs) in each FRG model, a tissue-specific parameter is computed and refined via a globally optimized scaling factor. Then, a two-hop-to-one-hop edge-promotion mechanism enhances graph connectivity without introducing notable heterophily. BARG is evaluated on a balanced training dataset of 1000 images and a test set of 554 TMA images (287 positive and 267 negative), with each patient contributing one image. Graph node features are extracted via a pre-trained VGG16 Convolutional Neural Network (CNN) by processing small image patches centered at nuclei detection peaks. Compared to the FRG approach, BARG yields notable performance gains, achieving 78% accuracy, 75% sensitivity, and 81% specificity, marking a 4% improvement in accuracy and an 8% increase in sensitivity. BARG also reaches an AUC-ROC of 0.85, a 3% enhancement over FRG, while preserving structural and contextual tissue relevance. These results position BARG as a robust, scalable solution for graph modelling in histopathology image analysis, suitable for broader applications in computational pathology.
Keywords: Breast Cancer Hormonal Status, Estrogen Receptor Status, Deep Learning, Graph Convolutional Networks, Digital Pathology, Machine Learning, Features Over-smoothing, Breast Cancer Biomarkers
Date Published: December 18, 2025 DOI: 10.11159/jbeb.2025.015
View Article