Augmented Multicenter Graph Convolutional Network for COVID-19 Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Graph Neural Networks

Background:

• Chest computed tomography (CT) scans for coronavirus 2019 (COVID-19) diagnosis often originate from diverse multicenter datasets with varying acquisition protocols.
• This inter-center heterogeneity poses a significant challenge for developing robust diagnostic models.
• Integrating data from multiple centers is crucial for increasing sample size and generalizability.

Purpose of the Study:

• To develop an effective method for diagnosing COVID-19 from multicenter chest CT scans.
• To address and mitigate the issue of inter-center heterogeneity in medical imaging datasets.
• To improve the accuracy and reliability of AI-driven COVID-19 diagnosis.

Main Methods:

• A 3-D convolutional neural network with a ghost module and multitask framework was used for initial feature extraction from CT scans.
• Extracted features were utilized to construct a multicenter graph, accounting for inter-center heterogeneity and disease status.
• An augmentation mechanism was employed to create an augmented multicenter graph for training the graph convolutional network (GCN).

Main Results:

• The proposed augmented multicenter graph convolutional network (AM-GCN) achieved a mean accuracy of 97.76% in diagnosing COVID-19.
• The model was validated on a large dataset comprising 2223 COVID-19 subjects and 2221 normal controls from seven medical centers.
• The developed method effectively handles data heterogeneity from different medical institutions.

Conclusions:

• The AM-GCN model demonstrates high performance and robustness in diagnosing COVID-19 using multicenter chest CT data.
• The approach successfully addresses the challenge of inter-center heterogeneity in medical imaging AI.
• Publicly available code facilitates further research and clinical application of this diagnostic tool.