Radiomic features based automatic classification of CT lung findings for COVID-19 patients

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• COVID-19 lung CT scans show Ground Glass Opacity (GGO), consolidation, and pleural effusion.
• GGOs often precede consolidations and exhibit heterogeneous appearances.
• Current severity scoring methods overlook the appearance of lung tissue, focusing only on the area of involvement.

Purpose of the Study:

• To identify heterogeneity/radiomic features that differentiate GGO, consolidation, and pleural effusion in COVID-19 lung CT images.
• To establish a baseline for feature selection in COVID-19 lung imaging analysis.
• To improve the accuracy of COVID-19 severity scoring.

Main Methods:

• Four feature selection approaches were evaluated from a pool of 44 features.
• Methods included manual selection, Genetic Algorithm-K-Nearest-Neighbor (GA-KNN), Genetic Algorithm-Binary Decision Tree (GA-BDT), and Genetic Algorithm-Artificial Neural Network (GA-ANN).
• An Artificial Neural Network (ANN) was trained and validated on independent datasets using selected features.

Main Results:

• Manual selection of nine radiomic features yielded the highest accuracy (85.7%) and AUC (0.90).
• Automatic methods (GA-BDT, GA-KNN, GA-ANN) showed lower accuracies (78%, 77.5%, 76.8%).
• Manually selected features demonstrated superior sensitivity and specificity.

Conclusions:

• Nine manually selected radiomic features enable accurate COVID-19 severity scoring.
• This method aids clinicians in planning personalized treatments and monitoring treatment response.
• The features are valuable for tracking COVID-19 progression and efficacy in clinical trials.