Deep Learning-Based Automatic CT Quantification of Coronavirus Disease 2019 Pneumonia: An International Collaborative...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Infectious Disease Diagnostics

Background:

• Accurate quantification of coronavirus disease 2019 (COVID-19) pneumonia on computed tomography (CT) images is crucial for patient management.
• Manual assessment of pneumonia extent and weight on CT scans can be time-consuming and subjective.
• Developing automated methods can improve efficiency and consistency in COVID-19 diagnosis and prognosis.

Purpose of the Study:

• To develop and validate an automated system for quantifying COVID-19 pneumonia on CT images.
• To assess the performance of the automated system against human-derived references.
• To evaluate the association of quantified pneumonia metrics with clinical symptoms and patient outcomes.

Main Methods:

• A retrospective study utilizing 176 chest CT scans from 131 COVID-19 patients across multiple institutions.
• Development of a 2D U-Net model for automatic segmentation of pneumonia based on radiologist-drawn masks.
• External validation using diverse datasets from Japan, Italy, China, and Radiopaedia, with performance measured by correlation coefficients.

Main Results:

• The automated system demonstrated high agreement with human references, with intraclass correlation coefficients of 0.990 for extent and 0.993 for weight in the internal dataset.
• Excellent correlation was observed in external validation datasets, with ICCs ranging from 0.949-0.965 for extent and 0.978-0.993 for weight.
• Quantified pneumonia extent and weight were independently associated with clinical symptoms and a composite outcome of respiratory failure and death.

Conclusions:

• Automated quantification of COVID-19 pneumonia on CT scans using the developed system is reliable and well-correlated with human assessments.
• The system's quantified metrics show independent associations with patient symptoms and prognosis.
• This automated approach offers a promising tool for objective assessment and outcome prediction in COVID-19 patients.