Automated Assessment and Tracking of COVID-19 Pulmonary Disease Severity on Chest Radiographs using Convolutional...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology
• Pulmonary Disease Assessment

Background:

• Chest radiographs (CXRs) are crucial for assessing COVID-19 pulmonary disease severity.
• Objective, automated tools are needed for consistent disease tracking and outcome prediction.

Purpose of the Study:

• To develop an automated measure of COVID-19 pulmonary disease severity on CXRs.
• To enable longitudinal disease tracking and predict patient outcomes.

Main Methods:

• A convolutional Siamese neural network was trained to generate a pulmonary x-ray severity (PXS) score.
• The algorithm utilized weakly-supervised pretraining and transfer learning on COVID-19 patient CXRs.
• Performance was evaluated on internal and external test sets, correlating PXS scores with radiologist assessments and predicting clinical outcomes.

Main Results:

• PXS scores demonstrated strong correlation with radiologist-assigned severity scores in both internal and external test sets (r=0.86).
• The PXS score change on follow-up CXRs agreed well with radiologist assessments (ρ=0.74).
• The PXS score effectively predicted subsequent intubation or death within three days for non-intubated patients (AUC=0.80).

Conclusions:

• A Siamese neural network-based PXS score provides an automated, reliable measure of COVID-19 pulmonary disease severity on CXRs.
• This automated score facilitates accurate disease tracking and aids in predicting critical patient outcomes, such as intubation or death.