A pyramid GNN model for CXR-based COVID-19 classification

The urgent need for efficient COVID-19 diagnosis has spurred advancements in chest X-ray (CXR) radiography, particularly with the aid of deep learning technologies like convolutional neural networks (CNNs) and graph neural networks (GNNs). Yet, the scarcity of labeled CXR images due to privacy constraints and the complexity of COVID-19 phenotypes often hamper model performance. In this study, we present an innovative pyramid GNN model that effectively tackles these challenges. By segmenting a CXR image into patches, our model leverages a CNN to capture shallow features, then employs a pyramid graph structure within GNN layers to gain the inter-relationship of infected region in distant patches and to amalgamate high-level features. These are subsequently processed by a multi-layer perceptron classifier for final diagnosis. Our approach offers multiple benefits, including noise elimination without the need for pre-treatment, efficient examination of remote infection regions, and the ability to accommodate the intricate structure of the lungs. Evaluations conducted on three distinct public CXR image datasets suggest that our pyramid GNN model offers a promising pathway for enhancing the accuracy and efficiency of COVID-19 diagnosis.