Any unique image biomarkers associated with COVID-19?

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Infectious Diseases

Background:

• Distinguishing COVID-19 from community-acquired pneumonia (CAP) using chest CT is crucial for patient management.
• Identifying unique imaging biomarkers for COVID-19 on CT scans is an ongoing area of research.

Purpose of the Study:

• To define unique chest CT infiltrative features associated with COVID-19.
• To evaluate the diagnostic performance of artificial intelligence (AI) models and human experts in differentiating COVID-19 from CAP on CT scans.

Main Methods:

• Retrospective collection of chest CT exams from 151 COVID-19 patients and 497 CAP patients.
• Development and testing of 3D convolutional neural network (CNN) classifiers to discriminate between COVID-19 and CAP.
• Visual interpretation of CT scans by experienced radiologists for comparison with AI model performance.

Main Results:

• AI models demonstrated moderate diagnostic ability (AUC 0.70) in distinguishing COVID-19 from CAP.
• The best-performing model identified 8-50% of CAP patients while misclassifying only 2% of COVID-19 patients.
• Radiologists' performance in visual interpretation was also assessed.

Conclusions:

• Both human experts and AI models have a moderately low ability to reliably distinguish all COVID-19 from CAP cases based on CT imaging alone.
• Automated image analysis shows promise for identifying a significant subset of non-COVID-19 cases, supporting targeted clinical decision-making.