Screening of COVID-19 Based on GLCM Features from CT Images Using Machine Learning Classifiers

Abstract

In healthcare, the decision-making process is crucial, including COVID-19 prevention methods should include fast diagnostic methods. Computed tomography (CT) is used to diagnose COVID patients' conditions. There is inherent variation in the texture of a CT image of COVID, much like the texture of a CT image of pneumonia. The process of diagnosing COVID images manually is difficult and challenging. Using low-resolution images and a small COVID dataset, the extraction of discriminant characteristics and fine-tuning of hyperparameters in classifiers provide challenges for computer-assisted diagnosis. In radiomics, quantitative image analysis is frequently used to evaluate the prognosis and diagnose diseases. This research tests an ML model built on GLCM features collected from chest CT images to screen for COVID-19. In this study, Support Vector Machines, K-nearest neighbors, Random Forest, and XGBoost classifiers are used together with LBGM. Tuning tests were used to regulate the hyperparameters of the model. With cross-validation, tenfold results were obtained. Random Forest and SVM were the best classification methods for GLCM features with an overall accuracy of 99.94%. The network's performance was assessed in terms of sensitivity, accuracy, and specificity.

Keywords: COVID-19; Feature extraction; GLCM; LGBM; Machine learning; SVM.

Fig. 1

Model of proposed method

Fig. 2

Sample images from the CT dataset. a Images of COVID-19 patients; b images of non-COVID-19 patients

Fig. 3

Calculation of GLCM

Fig. 4

Support vector machine

Fig. 5

Random forest tree

Fig. 6

LGBM-leafwise

Fig. 7

K-fold cross-validation (CV) model. P performance

Fig. 8

Confusion matrix for GLCM features. a SVM and b LBGM

Fig. 9

Confusion matrix for GLCM features. a RF, b KNN, and c XGB

Fig. 10

Performance of classifiers for GLCM features. a COVID and b non-COVID

Fig. 11

ROC curve