Deep learning analysis provides accurate COVID-19 diagnosis on chest computed tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonology

Background:

• Computed Tomography (CT) is crucial for COVID-19 diagnosis and management.
• High case-loads necessitate efficient, automated diagnostic tools.
• AI can expedite diagnosis and risk stratification for COVID-19.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) classifier for COVID-19 diagnosis using chest CT.
• To assess the diagnostic performance of the ML model against experienced radiologists.
• To establish validated rule-in and rule-out criteria for COVID-19 risk stratification.

Main Methods:

• A deep learning-derived ML classifier was developed using an open-source dataset of 6868 chest CT images.
• The model was trained and validated on subsets of the data.
• Performance was evaluated on an independent test set using Receiver Operating Characteristics (ROC) analysis and compared to radiologists.

Main Results:

• The ML model achieved an Area Under the Curve (AUC) of 0.956 on the independent test set.
• At the rule-in threshold, sensitivity was 84.4% and specificity was 93.3%, comparable to radiologists.
• At the rule-out threshold, sensitivity was 100% and specificity was 60%, differing significantly from radiologists.

Conclusions:

• A basic deep learning approach using open-source CT data can accurately diagnose COVID-19.
• The ML classifier provides validated criteria for risk stratification of COVID-19.
• This automated tool has the potential to significantly aid in COVID-19 diagnosis and management.