Table 4.
Overview of deep learning approaches for automated COVID-19 diagnosis.
| Reference | Task | Modality | Method | Total No. Of Images | No. Of Images From COVID-19 Cases | Accuracy (%) | Remarks |
|---|---|---|---|---|---|---|---|
| Wang and Wong (2020) [120] | Automatic COVID-19 diagnosis | CXR | CNN | 13,975 | 358 | 93.3 | COVID-Net has been proposed. |
| Narin et al. (2020) [113] | Automatic COVID-19 diagnosis | CXR | CNN | 100 | 50 | 98 | The pre-trained ResNet50 model provides the highest classification performance. |
| Hemdan et al. (2020) [114] | Automatic COVID-19 diagnosis | CXR | CNN | 75 | 25 | 90 | The VGG19 and DenseNet201 models showed a good and similar performance. |
| Ghoshal and Tucker (2020) [115] | Estimating uncertainty and interpretability in deep learning for COVID-19 diagnosis | CXR | BCNN | 5941 | 68 | 89 | Experiment has shown a strong correlation between model uncertainty and accuracy of prediction. |
| Apostolopoulos and Mpesiana (2020) [242] | Automatic COVID-19 diagnosis | CXR | CNN | 1442 | 224 | 96.78 | The MobileNet v2 effectively distinguished the COVID-19 cases from viral and bacterial pneumonia cases. |
| Apostolopoulos et al. (2020) [117] | Automatic classification of pulmonary diseases | CXR | CNN | 3905 | 455 | 99.18 | Mobile Net has been used for transfer learning. |
| Abbas et al. (2020) [243] | Automatic COVID-19 diagnosis | CXR | CNN | 196 | 105 | 95.12 | A deep CNN, called Decompose, Transfer, and Compose (DeTraC) has been validated. |
| Afshar et al. (2020) [244] | Automatic COVID-19 diagnosis | CXR | CNN | 13,975 | 358 | 95.7 | COVID-CAPS including several Capsule and convolutional layers has been proposed. |
| Chowdhury et al. (2020) [245] | Automatic COVID-19 diagnosis | CXR | CNN | 2876 | 190 | 98.3 | SqueezeNet outperforms AlexNet, ResNet18 and DenseNet201. |
| Oh et al. (2020) [121] | Automatic COVID-19 diagnosis | CXR | CNN | 15,043 | 180 | 91.9 | A patch-based deep neural network architecture that can be stably trained with small data set has been proposed. |
| Rajaraman et al. (2020) [246] | Automatic COVID-19 diagnosis | CXR | CNN | 16,700 | 313 | 99.01 | The best performing models are iteratively pruned to identify optimal number of neurons in the convolutional layers to reduce complexity and improve memory efficiency. |
| Luz et al. (2020) [247] | Automatic COVID-19 diagnosis | CXR | CNN | 13,800 | 183 | 93.9 | The proposed model has about 30 times parameters fewer than the baseline literature model, 28 and 5 times parameters fewer than the popular VGG16 and ResNet50 architectures, respectively. |
| Tartaglione et al. (2020) [248] | Automatic COVID-19 diagnosis | CXR | CNN | 584 | 405 | 95 | Possible obstacles in successfully training a deep model have been highlighted. |
| Hammoudi et al. (2020) [249] | Automatic COVID-19 diagnosis | CXR | CNN | 5863 | – | 95.72 | The DenseNet169 architecture has reached the best performance. |
| Khan et al. (2020) [250] | Automatic COVID-19 diagnosis | CXR | CNN | 1300 | 284 | 89.5 | CoroNet, a deep CNN based model, has been proposed. |
| Santosh et al. (2020) [251] | Automatic COVID-19 diagnosis | CXR | CNN | 6756 | 73 | 99.96 | The Truncated Inception Net deep learning model has been proposed. |
| Pereira et al. (2020) [252] | Automatic COVID-19 diagnosis | CXR | CNN | 1144 | 90 | – | A macro-avg F1-Score of 0.65 using a multi-class approach and an F1-Score of 0.89 for the COVID-19 identification in the hierarchical classification scenario have been achieved. |
| Murphy et al. (2020) [253] | Automatic COVID-19 diagnosis | CXR | CNN | 25,146 | 416 | – | An AUC of 0.81 has been achieved. The performance of an AI system in the detection of COVID-19 is comparable with that of six independent readers. |
| Ozturk et al. (2020) [254] | Automatic COVID-19 diagnosis | CXR | CNN | 1127 | 127 | 98.08 | The DarkCovidNet model has been proposed for binary and multi-class classification of COVID-19, no-Findings and pneumonia cases. |
| Togaçar et al. (2020) [255] | Automatic COVID-19 diagnosis | CXR | CNN | 458 | 295 | 99.27 | Features are extracted using deep learning architectures and classified by SVM. |
| Mahmud et al. (2020) [256] | Automatic COVID-19 diagnosis | CXR | CNN | 6161 | 305 | 97.4 | CovXNet architecture is proposed based on depthwise dilated convolutions. |
| Mahmoud et al. (2021) [257] | Automatic COVID-19 diagnosis | CXR | CNN | 15,496 | 589 | 95.82 | The CovidXrayNet model has been proposed for three-class classification. |
| Quan et al. (2021) [258] | Classification and segmentation of COVID-19 lesions | CXR | CNN | 9432 | 781 | 90.7 | The DenseCapsNet has been proposed. |
| Karakanis and Leontidis (2021) [259] | Automatic COVID-19 diagnosis | CXR | CNN and GAN | 435 | 145 | 98.7 | The GAN model has been used for data augmentation. |
| Jin et al. (2021) [260] | Automatic COVID-19 diagnosis | CXR | CNN | 1743 | 543 | 98.64 | A hybrid ensemble model, including a pre-trained AlexNet as feature extractor and an SVM classifier as the classifier, has been proposed. |
| Ahmad et al. (2021) [261] | Automatic COVID-19 diagnosis | CXR | CNN | 4000 | 1000 | 98.45 | Some of the existing CNN architectures with data augmentation have been used for COVID-19 diagnosis. |
| Zhang et al. (2021) [262] | Automatic COVID-19 diagnosis | CXR | CNN | 11,106 | 5806 | – | An AUC of 0.92 has been achieved for CV19-Net deep neural network architecture. The results show that the proposed method works better in diagnosing COVID-19 than experienced thoracic radiologists. |
| Wehbe et al. (2021) [263] | Automatic COVID-19 diagnosis | CXR | CNN | 14,002 | 5445 | 83 | The DeepCOVID-XR architecture shows similar performance to experienced thoracic radiologists. |
| Keidar et al. (2021) [264] | Automatic COVID-19 diagnosis | CXR | CNN | 2426 | 1289 | 90.3 | Some pre-trained deep CNN architectures with data augmentation have been used. |
| Li et al. (2020) [265] | Automatic COVID-19 diagnosis | CT | CNN | 4356 | 1296 | – | An AUC of 0.96 for detecting COVID-19 has been achieved. |
| Huang et al. (2020) [266] | Evaluation of lung burden changes in patients with COVID-19 | CT | CNN | 126 | 126 | – | A commercially available deep-learning-based tool has been used. |
| Zheng et al. (2020) [118] | Automatic COVID-19 diagnosis | CT | CNN | 630 | – | 90.1 | A pre-trained U-Net for lung segmentation and a 3D CNN architecture (DeCoVNet) have been used. |
| Chen et al. (2020) [267] | Automatic COVID-19 diagnosis | CT | CNN | 35,355 | 20,886 | 95.24 | U-NET++ has been used for retrospective and prospective COVID-19 dataset evaluation. |
| Hu et al. (2020) [119] | Automatic COVID-19 diagnosis | CT | CNN | 450 | 150 | 96.2 | A weakly-supervised deep learning framework for fast and fully-automated detection and classification of COVID-19 has been presented. |
| Loey et al. (2020) [116] | Automatic COVID-19 diagnosis | CT | CNN and CGAN | 742 | 345 | 82.91 | Data augmentations along with CGAN improve the performance of classification in AlexNet, VGGNet16, VGGNet19, GoogleNet, and ResNet50 deep transfer models. |
| Wu et al. (2020) [268] | Classification and Segmentation for COVID-19 diagnosis | CT | CNN | 144,167 | 68,626 | – | A Joint Classification and Segmentation (JCS) system obtains an average sensitivity of 95.0% and a specificity of 93.0% on the classification test set, and 78.3% Dice score on the segmentation test set. |
| Li et al. (2020) [269] | Automatic COVID-19 diagnosis | CT | CNN | 4352 | 1292 | – | The sensitivity and specificity for detecting COVID-19 are 90% and 96% respectively, with an AUC of 0.96. |
| Bai et al. (2020) [270] | Differentiating COVID-19 and other pneumonia | CT | CNN | 132,583 | – | 96 | Artificial intelligence improved radiologists' performance in distinguishing COVID-19 from other pneumonia. |
| Pu et al. (2020) [271] | Automatic COVID-19 diagnosis | CT | CNN | 955 | 498 | – | An AUC of 0.70 has been achieved. |
| Ni et al. (2020) [272] | Automatic COVID-19 diagnosis | CT | CNN | 19,291 | 3854 | 94 | The deep learning model improves diagnosis efficiency by shortening processing time. |
| Li et al. (2020) [273] | Segmentation of COVID-19 chest CT images | CT | CNN | 558 | 558 | – | The dice coefficient between the proposed method's segmentation and two experienced radiologists for the COVID-19-infected lung abnormalities is 0.74 and 0.76, respectively. |
| Ardakani et al. (2020) [274] | Automatic COVID-19 diagnosis | CT | CNN | 1020 | 510 | 99.63 | Different well-known CNN architectures were evaluated for COVID-19 diagnosis. ResNet-101 and Xception show the best performance. |
| Amyar et al. (2020) [275] | Classification and segmentation of COVID-19 lesions | CT | AE | 1369 | 449 | 94.67 | The dice coefficient of 88% was obtained using multi-task deep learning based model for image segmentation. |
| Serte and Demirel (2021) [276] | Automatic COVID-19 diagnosis | CT | CNN | 7572 | 2496 | 98 | The proposed method combined the ResNet-50 model and the majority voting with an AUC of 96% as the best result. |
| Arora et al. (2021) [277] | Automatic COVID-19 diagnosis | CT | CNN | 3294 | 1601 | 100 | Some of the pre-trained deep models have been evaluated for COVID-19 diagnosis using CT images. |
| Zhao et al. (2021) [278] | Segmentation of COVID-19 lesions | CT | CNN | 2317 | 2317 | – | A dilated dual attention U-Net based on the dual attention strategy and hybrid dilated convolutions has been proposed for COVID-19 lesion segmentation in CT images. A Dice score of 0.72 has been achieved. |
| Maghdid et al. (2020) [279] | Automatic COVID-19 diagnosis | CXR and CT | CNN | CXR: 170 CT: 361 |
CXR: 85 CT: 203 |
98 | The utilised models can provide accuracy up to 98% via pre-trained AlexNet and 94.1% accuracy by using the modified CNN. |
| Jia et al. (2021) [280] | Automatic COVID-19 diagnosis | CXR and CT | CNN | CXR: 7592 CT: 104,009 |
CXR: 1770 CT: Not clear |
CXR: 99.6 CT: 99.3 |
The modified MobileNet and ResNet have been proposed. |
| Chaudhary and Pachori (2021) [281] | Automatic COVID-19 diagnosis | CXR and CT | CNN | CXR: 1446 CT: 2481 |
CXR: 482 CT: 1252 |
CXR: 100 CT: 97.6 |
The combination of Fourier-Bessel series expansion-based image decomposition, different CNN architectures and various classifiers have been evaluated. |
| Ibrahim et al. (2021) [282] | Automatic COVID-19 diagnosis | CXR and CT | CNN and GRU | 33,676 | 4320 | 98.05 | A multi-class classification method including VGG19 and some additional CNN layers shows the best performance. |