Table 1.
Comparative summary of the proposed and the existing state-of-the-art methods for automatic diagnosis of COVID19 pneumonia. (V.C: Varicella, S.T: Streptococcus, P.M: Pneumocystis, T.B: Tuberculosis, M.R: MERS, S.R: SARS, P.N: Pneumonia, B.P: Bacterial pneumonia, V.P: Viral pneumonia, ACC: Accuracy, F1: F1 score, AP: Average precision, AR: Average recall, Spec: Specificity, Kap: Kappa statistics, Others: Combination of different diseased and normal classes beside COVID19).
| Literature | Method | Class name (No. of classes) | Dataset (No. of images) |
Imaging modality | Result (%) | Strength | Limitation | |
|---|---|---|---|---|---|---|---|---|
| COVID19＋ | COVID19 | |||||||
| Ghoshal et al. [6] | Bayesian CNN | COVID19/B.P/V.P/ Normal (4) |
68 | 5873 | X-ray | ACC: 89.82 | Enhanced detection performance compared to Standard ResNet | Limited COVID19 data samples |
| Oh et al. [7] |
FC-DenseNet+ResNet18 | COVID19/B.P/V.P/T.B/ Normal (5) |
180 | 322 | X-ray | ACC: 88.9, F1: 84.4, AP: 83.4, AR: 85.9, Spec: 96.4 | Provides clinically interpretable saliency maps | - Limited COVID19 data samples - Patch processing required high computational cost |
| Das et al. [8] |
Truncated InceptionNet | COVID19/P.N/T.B/ Normal (4) |
162 | 6683 | X-ray | ACC: 98.7, F1: 97, AP: 99, AR: 95, Spec: 99, AUC: 99 | Enhanced computational performance compared to Standard InceptionNet | Limited COVID19 data samples |
| Singh et al. [9] | MODE-based CNN | COVID19/Others (2) |
69 | 63 | CT | ACC: 93.5, F1: 89.9, AR: 90.75, Spec: 90.8, Kap: 90.5 | Applicable in real-time screening | - Limited dataset - Lack of ablation study |
| Pereira et al. [10] | Texture Descriptors and InceptionNet | COVID19/M.R/S.R/ V.C/S.T/P.M/Normal (7) |
180 | 2108 | X-ray | F1: 89 | High COVID19 recognition rate | - Limited COVID19 data samples - Required high computational power |
| Khan et al. [11] |
CoroNet | COVID19/B.P/V.P/ Normal (4) |
284 | 967 | X-ray | ACC: 89.6, F1: 89.8, AP: 90, AR: 89.92, Spec: 96.4 | High COVID19 detection rate compared to other classes | - Limited COVID19 data samples - Lack of ablation study |
| Asnaoui et al. [12] | InceptionResNet | COVID19/B.P/Normal (3) |
231 | 5856 | X-ray | F1: 92.08, AP: 92.38, AR: 92.11, Spec: 96.06 | Detailed performance analysis under the same experimental protocol | - Low COVID19 detection rate compared to other classes - Limited COVID19 data samples |
| Brunese et al. [13] |
VGG16 | COVID19/P.N/Normal (3) |
250 | 6273 | X-ray | ACC: 97 | High COVID19 detection rate compared to other classes | - Lack of ablation study - Limited COVID19 data samples |
| Han et al. [14] | AD3D-MIL | COVID19/P.N/Normal (3) |
230 | 230 | CT | ACC: 94.3, F1: 92.3, AP: 95.9, AR: 90.5, AUC: 98.8, Kap: 91.1 | Provides clinically interpretable saliency maps | - Required high computational power |
| Mahmud et al. [15] | CovXNet | COVID19/B.P/V.P/ Normal (4) |
305 | 915 | X-ray | ACC: 90.2, F1: 90.4, AP: 90.8, AR: 89.9, Spec: 89.1, AUC: 91 | Generate clinically interpretable activation maps | - Limited dataset - Low COVID19 detection rate |
| Proposed | Ensemble-Net | COVID19/Others (2) | 3296 | 4143 | X-ray | ACC: 95.83, F1: 95.94, AP: 95.68, AR: 96.20, AUC: 97.99 | - Reduced number of trainable parameters - Visualize clinically interpretable ML-CAMs | - Training time is longer than the conventional end-to-end training method |
| 3254 | 2217 | CT | ACC: 94.72, F1: 94.60, AP: 95.22, AR: 94.00, AUC: 97.5 | |||||