Effective Estrogen Receptor Status Classification in BreastCancer Using a Single TMA Image and HGCN-CNN Ensemble

Introduction: This study proposes a Computer-Aided Diagnosis (CAD) method to determine Estrogen Receptor Status (ERS) in breast cancer patients using Hematoxylin and Eosin (H&E)-stained tissue microarray (TMA) images. It is the first to explore the contributions of graph-based analysis for this task and their integration with conventional Convolutional Neural Network (CNN)-based approaches. Methods: A graph model of H&E-stained images is proposed, representing intercellular spatial dependencies as graph edges. We explore the significance of intercellular spatial relationship properties derived from cellular morphology and graph topology, which have been unexplored in previous studies. Additionally, this work combines the results of the proposed graph classifier with those of well-established Convolutional Neural Network(CNN) classifiers to optimize accuracy. Results: The proposed Hybrid Graph Convolutional Network (HGCN) achieves 14%,25%, and 39% higher specificity compared to the popular pre-trained CNN classifiers VGG16, DenseNet121, and ResNet50, respectively. Whereas, HGCN demonstrates lower sensitivity than these CNN classifiers. By combining the strengths of both approaches in a novel HGCN-CNN ensemble, a well-balanced trade-off between sensitivity and specificity is achieved, resulting in the highest Area Under the Receiver Operating CharacteristicCurve (AUC-ROC) of 87% among all the classifiers. The ensemble outperforms individual CNN classifiers in accuracy, AUC-ROC, and specificity while also exceeding the HGCN classifier in accuracy, AUC-ROC, and sensitivity. These results are obtained from a training dataset of 1,000 H&E-stained TMA images and a test dataset of 554 TMA images. Conclusions: The results demonstrate that the graph-based approach offers significantly higher specificity compared to popular pre-trained CNN classifiers, while the CNN classifiers excel in sensitivity. The proposed HGCN-CNN ensemble effectively leverages the complementary strengths of both classifier types, enhancing overall binary classification performance in terms of AUC-ROC and accuracy compared to all individual classifiers.