[Paper Review] Covid-19 Detection from Chest X-ray and Patient Metadata using Graph Convolutional Neural Networks

Thosini Bamunu Mudiyanselage, Nipuna Senanayake+3 more|arXiv (Cornell University)|May 20, 2021

COVID-19 diagnosis using AI16 references4 citations

TL;DR

This paper proposes a Graph Convolutional Neural Network (GCN) that integrates chest X-ray (CXR) images and patient metadata to detect COVID-19, outperforming transfer learning models like ResNet-50. By modeling relationships between data instances via a similarity-based graph and leveraging a CNN encoder for feature extraction, the GCN achieves 94% accuracy in binary classification and 86% in three-class classification, with superior generalization and reduced false positives.

ABSTRACT

The novel corona virus (Covid-19) has introduced significant challenges due to its rapid spreading nature through respiratory transmission. As a result, there is a huge demand for Artificial Intelligence (AI) based quick disease diagnosis methods as an alternative to high demand tests such as Polymerase Chain Reaction (PCR). Chest X-ray (CXR) Image analysis is such cost-effective radiography technique due to resource availability and quick screening. But, a sufficient and systematic data collection that is required by complex deep leaning (DL) models is more difficult and hence there are recent efforts that utilize transfer learning to address this issue. Still these transfer learnt models suffer from lack of generalization and increased bias to the training dataset resulting poor performance for unseen data. Limited correlation of the transferred features from the pre-trained model to a specific medical imaging domain like X-ray and overfitting on fewer data can be reasons for this circumstance. In this work, we propose a novel Graph Convolution Neural Network (GCN) that is capable of identifying bio-markers of Covid-19 pneumonia from CXR images and meta information about patients. The proposed method exploits important relational knowledge between data instances and their features using graph representation and applies convolution to learn the graph data which is not possible with conventional convolution on Euclidean domain. The results of extensive experiments of proposed model on binary (Covid vs normal) and three class (Covid, normal, other pneumonia) classification problems outperform different benchmark transfer learnt models, hence overcoming the aforementioned drawbacks.

Motivation & Objective

• To address the limitations of transfer learning models in medical imaging, such as poor generalization and high bias due to limited, domain-specific data.
• To improve detection of COVID-19 in CXR images by integrating both imaging features and patient metadata through a relational graph structure.
• To develop a model that generalizes better than existing transfer learning approaches on small, imbalanced medical datasets.
• To enable accurate differentiation between COVID-19, normal, and other pneumonia types in a multi-class classification setting.
• To reduce false positives in diagnosis by leveraging graph-based relational learning and attention-aware feature extraction.

Proposed method

• A CNN encoder (e.g., VGG16-like architecture) is used to extract deep features from CXR images, which are then embedded into a graph structure based on instance similarity.
• A graph is constructed where nodes represent data samples (CXR images with metadata), and edges are weighted by cosine similarity between feature vectors.
• Graph Convolutional Networks (GCNs) apply message-passing operations using mean aggregation to propagate and update node representations across the graph.
• The GCN uses a softmax layer on the final graph-embedded features for multi-class classification into normal, COVID-19, bacterial, viral, and fungal pneumonia.
• Saliency maps are generated to visualize attention regions in CXR images, validating that the model focuses on clinically relevant opacities.
• The model is trained end-to-end with cross-entropy loss and evaluated using 5-fold cross-validation on a real-world CXR dataset.

Experimental results

Research questions

• RQ1 Can a GCN-based model outperform conventional transfer learning models like ResNet-50 in classifying COVID-19 from CXR images?
• RQ2 Does incorporating patient metadata and relational knowledge between samples improve model generalization and reduce overfitting?
• RQ3 How does the choice of graph aggregation function (mean, sum, max) affect performance on medical graph data?
• RQ4 Can the model effectively distinguish between COVID-19 and other types of pneumonia in a multi-class setting?
• RQ5 To what extent does the CNN encoder’s attention distribution correlate with radiological findings in CXR images?

Key findings

• The proposed GCN achieved 94% accuracy in binary classification (COVID-19 vs. normal), outperforming ResNet-50, InceptionV3, and VGG16.
• In three-class classification (COVID-19, normal, other pneumonia), the GCN achieved 86% accuracy, demonstrating strong performance on complex diagnostic tasks.
• The mean aggregation function in GCN showed better generalization than sum or max pooling, especially in denser graphs.
• Confusion matrix analysis revealed that the GCN had significantly fewer false positives—especially for viral, bacterial, and fungal pneumonia misclassified as COVID-19—compared to ResNet-50.
• Saliency maps confirmed that the GCN’s CNN encoder focused attention on clinically relevant opacities in CXR images, unlike VGG16 and InceptionV3, which showed scattered or misdirected attention.
• The GCN model demonstrated superior performance in identifying true positive cases across all classes, particularly in detecting COVID-19 and bacterial pneumonia, which is critical for controlling disease spread.