Multi-classifier-based identification of COVID-19 from chest computed tomography using generalizable and...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Distinguishing coronavirus disease (COVID-19) from other viral pneumonias based on clinical and CT findings can be challenging.
• Radiomics offers a quantitative approach to analyze medical images, potentially improving diagnostic accuracy.

Purpose of the Study:

• To evaluate the effectiveness of radiomics in diagnosing COVID-19 compared to other viral pneumonias with similar presentations.
• To identify specific radiomic features that can differentiate between COVID-19 and other viral pneumonia cases.

Main Methods:

• Retrospective analysis of CT scans from 110 COVID-19 positive and 108 COVID-19 negative patients.
• Manual segmentation of pneumonia lesions and extraction of 120 radiomic features using Pyradiomics.
• Application of four classifiers (linear classifier, k-nearest neighbour, LASSO, random forest) for differentiation and comparison with radiologist performance.

Main Results:

• The Least Absolute Shrinkage and Selection Operator (LASSO) classifier achieved the best performance (AUC: 0.81) in differentiating COVID-19.
• Radiomic analysis showed excellent agreement (Kappa score: 0.89) with radiologists' assessments.
• Specific radiomic features like "Original_Firstorder_RootMeanSquared" and "Original_Firstorder_Uniformity" were identified as significant for classification.

Conclusions:

• Radiomics, utilizing specific features, can significantly aid in classifying COVID-19 pneumonia.
• This quantitative approach enhances the understanding of CT imaging characteristics differentiating COVID-19 from other viral pneumonias.