Diagnosis of COVID-19 Pneumonia Based on Graph Convolutional Network

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Chest computed tomography (CT) datasets exhibit diversity in equipment types, posing challenges for automated analysis.
• Accurate and rapid diagnosis of coronavirus disease 2019 (COVID-19) is crucial for patient management and reducing healthcare burdens.

Purpose of the Study:

• To develop a 3D deep learning method for rapid and accurate COVID-19 diagnosis from chest CT images.
• To address dataset diversity by proposing a graph-based approach within a graph convolutional network (GCN).

Main Methods:

• A 3D convolutional neural network (3D-CNN) with transfer learning was used to extract image features.
• A COVID-19 graph was constructed using extracted features, clustering samples by equipment type.
• Edge weights were computed by combining feature correlation distance and 3D-CNN score differences.
• The COVID-19 graph was input into a GCN for final diagnosis.

Main Results:

• The method achieved high diagnostic performance on a dataset of 399 COVID-19 cases and 400 normal controls.
• Accuracy reached 98.5%, sensitivity 99.9%, and specificity 97% across six equipment types.

Conclusions:

• The proposed 3D deep learning and GCN method offers a robust solution for COVID-19 diagnosis from diverse CT datasets.
• This approach can significantly reduce the workload for radiologists and physicians in diagnosing COVID-19.