Fig 1.
(a) Classifying COVID-19 pneumonia from non-COVID-19 cases: COVID-19 infected, other pneumonia infected, and uninfected CT images were classified using three different pre-trained models VGG16, VGG19 and Xception, with ML classifiers, SVM, Random Forest, Logistic regression, and XGBoost (b) Prediction of cardiac abnormalities in COVID-19 infected patients: 3D-CNN model consists of three convolution layers followed by ReLU, Maxpool layer, batch normalization layer for feature extraction from the cardiac region, with the classification part consists of one dense layer, ReLU, dropout layer, and activation function sigmoid. Finally, it diagnosed abnormal cardiac health from CT-scan of COVID-19 infected and healthy subjects. (c) Classification of cardiovascular abnormalities using CT-measured cardiac parameters. Cardiac parameters were extracted from the 3D CT slices of COVID-19 infected patients and uninfected patients. The parameters were 1) cardiac and thoracic diameter ratio, 2) pulmonary artery and aorta ratio, 3) presence of calcified plaque and 4) mean epicardial adipose tissue area. ML classifier support vector machine (SVM) was then used to predict the impact of COVID-19 on cardiac health from CT measured parameters.
Fig 2.
Network architecture for the proposed method.
Initially, pre-trained model VGGNet has been used as the most significant CNN base deep feature extraction technique, which is subsequently fed into supervised ML models for the final representation of a proposed hybrid model.
Fig 3.
Flowchart of proposed 3D-CNN model applied to detect COVID-19 associated cardiac abnormalities from CT scan images.
Fig 4.
(a) Confusion matrices and (b) ROC curve for multiclass classification with pre-trained model VGG19 and machine learning classifiers SVM, showing the performance of classification on independent test data set. Images classified into three classes of COVID-19 pneumonia, other pneumonia, and healthy subjects.
Table 1.
Computational time and execution time for pretrained model and proposed hybrid model.
Fig 5.
(a) Training and validation accuracy curve, and loss curve as achieved with augmented training data sets with 97% accuracy. (b) Confusion matrix and (c) ROC curve to evaluate the accuracy after 3D classification of the cardiac component from the CT scans of COVID-19 infected and uninfected healthy subjects.
Fig 6.
(a) Percentage of cardiac and thoracic diameter ratio, (b) pulmonary artery and aorta ratio; (c) average adipose tissue area and (d) percentage of calcium plaque are plotted for 162 COVID-19 positives and 167 COVID-19 negative patients. The percentage of each case is higher in positive cases than negative cases. The datasets were found to be highly statistically significant among each other: P < 0.0001 (***). (e) confusion matrix and (f) ROC-AUC curve for 10-fold cross validation performed on cardiac parameters.