Robust framework for COVID-19 identication from a multicenter dataset of chest CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate differentiation of COVID-19, Community-Acquired Pneumonia (CAP), and normal cases from chest CT scans is crucial for patient management.
• Deep learning models often struggle with data heterogeneity from different imaging centers and scanners.

Purpose of the Study:

• To develop a robust deep learning framework for classifying COVID-19, CAP, and normal cases using volumetric chest CT scans.
• To enhance the model's adaptability and performance on heterogeneous datasets through unsupervised learning and ensemble methods.

Main Methods:

• A deep learning model was trained on chest CT scans from a single center.
• The model was evaluated on diverse test sets from multiple scanners and protocols, including low-dose CT.
• An unsupervised approach was used to update the model with confident predictions from test data, and an ensemble architecture aggregated predictions.

Main Results:

• The framework achieved high accuracy (96.15%) and sensitivity for COVID-19 (96.08%), CAP (92.86%), and normal cases (98.04%) on heterogeneous test sets.
• Area Under the Curve (AUC) values exceeded 0.989 for all classes.
• The unsupervised enhancement significantly improved model performance and robustness on varied external datasets.

Conclusions:

• The proposed deep learning framework demonstrates robust performance in classifying COVID-19, CAP, and normal cases from diverse chest CT scans.
• Unsupervised model updating and ensemble methods effectively address data shift and improve generalization capabilities.
• This approach offers a promising tool for reliable COVID-19 diagnosis in varied clinical settings.