Table 5B.
Objectives, results, advantages and disadvantages of the included novel COVID-19-AI imaging studies.
| Experimental study | Research direction & AI technology | Results | Advantages | Disadvantages |
|---|---|---|---|---|
| (Zhang et al., 2020a, 2020b) | Triage of COVID-19 pneumonia using the deep learning software uAI. | The uAI Intelligent Assistant Analysis System detected COVID-19 Pneumonia in addition to COVID-19 CT findings (i.e. GGOs and lobular lesions) | Useful for localization and quantification of lung lesions | Requires manual human modification to rollout negative lesions |
| Useful for treatment planning based on the affected lung regions | ||||
| Not applicable for critical cases (CT) | ||||
| (Chowdhury et al., 2020) | Autonomous detection of COVID-19 pneumonia using transfer learning. | The transfer learning system managed to train CNNs to achieve readings of 99.7 %, 99.7 %, 99.7 %, 99.55 %,97.9 %, 97.95 %, 97.9 %, and 98.8 % in taxonomical accuracy, pathological sensitivity, and specificity to COVID-19. | Useful for normal and abnormal classification Image augmentation can slightly enhance the overall performance | Large data set is required to improve the accuracy |
| (Zhu et al., 2020) | Classification of COVID-19 severity on portable CXRs using deep learning CNNs and an expert radiologist panel. | The deep learning CNNs accomplished a comparable staging accuracy to the three-member radiologist panel at a mean absolute error of 8.5 %. | Incorporating AI with portable X-ray images provide more accessible diagnosis than CT scan | Small data set was used |
| No Correlation with radiographic score system | ||||
| No correlation with clinical non imaging information | ||||
| Useful for disease severity identification | ||||
| Transfer learning is superior to traditional learning techniques in terms of shorter training time | ||||
| (Ozturk et al., 2020) | Autonomous detection of COVID-19 using deep learning CNNs and an expert radiologist panel to compare normal Vs. COVID-19 and COVID-19 Vs. pneumonia. | The proposed obtained an accuracy of 98.08 % in binary classification and 87.02 % in multi-class classification. | No manual extraction is needed | Small data set was used |
| Incorporating AI with X-ray images provide more accessible diagnosis than CT scan Heat-maps images are useful for localization | ||||
| (Hwang et al., 2020) | Evaluation of deep learning CAD (computer-aided detection) system performance in image interpretation of suspected COVID-19 CXRs. | AI CAD system generated a 68.8% sensitivity and 66.7% specificity to COVID-19. COVID-19 pneumonia was also detected at a 72.3% specificity and 81.5% sensitivity. | CAD technique provides shorter diagnostic time than PCR result | Lack of trained data on COVID-19 images |
| (Bai et al., 2020) | Evaluation of AI integrated image interpretation workflow in the differentiation of COVID-19 and other pulmonary findings on chest CTs. | Deep learning apparatus assisted radiologists in improving the diagnostic performance in terms of COVID-19 at a 90% accuracy, 91% specificity, and 88% sensitivity. | AI augmentation is useful for differentiating COVID-19 from other pneumonia on CT images | Small data set was used |
| Lack of homogenous pneumonia cases | ||||
| (Apostolopoulos et al., 2020) | Autonomous detection of COVID-19 using transfer learning, deep learning, and CNNs. | The proposed study method delivered 99.42 % specificity, 99.18 % accuracy, and 97.36 % sensitivity in the identification of biological markers of COVID-19. | Useful to identify new pulmonary abnormalities as new biomarkers | Small data set was used |
| Lack of suspected COVID-19 patients data | ||||
| (Tsiknakis et al., 2020) | Development of a feasible AI model in terms of image interpretation of COVID-19 CXRs using transfer learning techniques and the evaluation of an expert radiologist panel. | The transfer learning model was capable of undertaking binary, ternary, and quaternary at the area under curve of 1 during the management of a 5 stage dataset. | Attention maps can improve the final clinical decision | limited data set |
| (Ko et al., 2020) | Rapid triage, investigation, and differentiation of the pulmonology of COVID-19, pneumonia, and non-pneumonia disorders of the lungs using transfer learning. | The ResNet-50 model outperformed the transfer learning models at a 99.87 % accuracy, 100 % specificity, and a 99.58 % sensitivity. Hence, providing a reliable diagnostic detection of COVID-19 pneumonia. | Useful for differentiating COVID-19 from other pneumonia on CT images | Small data set was used Lack of data for validation |
| (Song et al., 2020) | Autonomous differentiation of viral pneumonia from COVID-19 computed tomography findings using bi-directional generative adversarial network data architecture to enhance unsupervised learning of the presented data. | The novel AI platform generated a maximum specificity of 91 % and a sensitivity of 92 % during the training, testing, and validation stages of the study. | Useful for differentiating COVID-19 from other pneumonia on CT images | Small data set was used Variation in CT image acquisition parameters Lack of data of severe COVID-19 cases |
| (Ni et al., 2020) | Autonomous recognition of CXR COVID-19 findings using deep learning. | The deep learning platform maintained a consistent diagnostic performance in comparison to the expert radiologists at a maximum accuracy of 0.85 Vs. 0.93 of expert medical imaging residents. | Useful for quantitative detection of abnormalities | Small data set was used |
| Lack of specificity in comparison to radiologists | ||||
| (Ran et al., 2020) | Autonomous differentiation of COVID-19 pneumonia from other causes of CXR abnormalities and test the system performance against thoracic radiologists using deep neural networks. | CV19-Net was able to differentiate COVID-19 related pneumonia from other types of pneumonia with performance paralleling that of experienced thoracic radiologists at a confidence interval of 95%. The combination of chest radiography with the proposed CV19-Net deep learning algorithm has the potential as an accurate method to improve the accuracy and the estimated times of the radiological interpretation process of COVID-19 pneumonia. | Useful for differentiating COVID-19 from other pneumonia on CT images | Lack of pneumonia types classification |
| Lack of COVID-19 representative data |