The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, which are based on the Cochrane Handbook for Systematic Reviews of Interventions, were used to conduct this study [15].

We searched electronic databases, such as PubMed, Scopus, Google Scholar, and Web of Science, for articles published between January 1 and April 25, 2020. We developed a search strategy using combinations of the following Medical Subject Headings: “COVID-19,” or “coronavirus,” or “SARS-CoV-2,” or “novel corona,” or “2019-ncov,” and “deep learning,” or “artificial intelligence,” or “automatic detection.” Reference lists of the retrieved articles and relevant reviews were also checked for additional eligible articles.

During the first screening, two authors (MMI and TNP) assessed the title and abstract of each article and excluded irrelevant articles. To include eligible articles, those two authors examined the full text of the articles and evaluated whether they fulfilled the inclusion criteria of this study. Disagreement during this selection process was resolved by consensus or, if necessary, the main investigator (YCL) was consulted. We included articles if they met the following criteria: (1) were published in English, (2) were published in a peer-reviewed journal, (3) assessed performance of a DL model to detect COVID-19, and (4) provided a clear description of the methodology and the total number of images. We excluded studies if they were published in preprint repositories or if they were published in the form of a review or a letter to the editor.

Two authors (MMI and TNP) independently screened all titles and abstracts of retrieved articles. The most relevant studies were selected based on the predefined selection criteria. Any disagreement during the screening process was resolved by discussion with the other authors; unsettled issues were settled by discussion with the study supervisor (YCL). The two authors who conducted the first screening cross-checked studies for duplication by comparing author names, publication dates, and journal names. They excluded all duplicate studies. Afterward, they collected data from the selected studies, such as author name, publication year, location, model description, total number of images, total number of COVID-19 cases and images, imaging modality, total number of patients, sensitivity, specificity, accuracy, area under the receiver operating characteristic curve (AUROC), and database.

The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool was used to assess the quality of the selected studies [16]. The QUADAS-2 scale comprises four domains: patient selection, index test, reference standard, and flow and timing. The first three domains are used to evaluate the risk of bias in terms of concerns regarding applicability. The overall risk of bias was categorized into three groups: low, high, and unclear risk of bias.

Meta-DiSc, version 1.4, was used to calculate the evaluation metrics of the DL model. The software was also used to (1) perform statistical pooling from each study and (2) assess the homogeneity with a variety of statistics, including chi-square and I². The sensitivity and specificity with 95% CIs in distinguishing between COVID-19 patients, patients with other types of pneumonia, and normal patients were calculated. The pooled receiver operating characteristic (ROC) curve was plotted and the area under the curve (AUC) was calculated with 95% CIs based on the DerSimonian-Laird random effects model method. The diagnostic odds ratio (DOR) was calculated by the Moses constant of the linear model. Diagnostic tests where the DOR is constant, regardless of the diagnostic threshold, have symmetrical curves around the sensitivity-specificity line. In these situations, it is possible to combine DORs using the DerSimonian-Laird method to estimate the overall DOR and, hence, to determine the best-fitting ROC curve [17]. The mathematical equation is given below:

When the DOR changes with the diagnostic threshold, the ROC curve is asymmetrical. To fit the DOR variation based on a different threshold, the Moses-Shapiro-Littenberg method was used. It consists of observing the relationship by fitting the straight line:

where D is the log of DOR and S is a measure of threshold given by the following:

Estimates of parameters a and b and their standard errors and covariance were obtained by the ordinary or weighted least squares method using the NAG Library for C (The Numerical Algorithms Group).

The ROC curve is the AUC that summarized the diagnostic performance as a single number: an AUC close to 1 is considered a perfect curve and an AUC close to 0.5 is considered poor [18]. The AUC is computed by numeric integration of the curve equation by the trapezoidal method [19]. The Q* index is defined by the point where sensitivity and specificity are equal, which is the point closest to the ideal top-left corner of the ROC curve space. It was calculated by the following:

Moreover, the standard error of the AUROC was calculated by following equation:

The standard error of Q* was calculated by following equation:

Figure 1 shows the process of identifying relevant DL studies. A total of 562 studies were retrieved by searching electronic databases and by reviewing their reference lists. We excluded 435 duplicate studies and an additional 104 studies that did not fulfill the selection criteria. We reviewed 23 full-text studies and further excluded 7 studies because of the reasons shown in Figure 1. Finally, we included 16 studies in the meta-analysis [13,14,20-33].

Among the 16 DL-based COVID-19 detection studies, we identified 5896 digital images for COVID-19 patients and 645,825 images for non-COVID-19 patients, including those with other types of viral pneumonia and normal patients. Included studies used DL algorithms, such as convolutional neural networks, MobileNetV2, and COVNet, for stratifying COVID-19 patients with higher accuracy. The range of accuracy for detecting COVID-19 correctly was 76.00 to 99.51. A total of 8 studies used CT images and 8 studies used x-ray images. The characteristics of the included studies in the meta-analysis are shown in Table 1 [13,14,20-33].

Based on the 16 studies, the performance of the DL algorithms for detecting COVID-19 was determined and is summarized in Table 2 [22,24,33]. The pooled sensitivity and specificity of the DL methods for detecting COVID-19 was 0.95 (95% CI 0.94-0.95) and 0.96 (95% CI 0.96-0.97), respectively, with a summary ROC (SROC) of 0.98 (Figure 2). The pooled sensitivity and specificity are shown in Figure 3.

DL methods were able to correctly distinguish other types of pneumonia from COVID-19 with an SROC of 0.98 (sensitivity: 0.93, 95% CI 0.92-0.94; specificity: 0.95, 95% CI 0.94-0.95). The positive likelihood, negative likelihood, and DOR were 22.45 (95% CI 12.86-39.19), 0.06 (95% CI 0.03-0.13), and 461.81 (95% CI 134.96-1580.24), respectively. Moreover, the DL model showed good performance for correctly stratifying normal patients, with an SROC of 0.99 (sensitivity: 0.95, 95% CI 0.94-0.96; specificity: 0.98, 95% CI 0.97-0.98). The positive likelihood, negative likelihood, and DOR were 47.47 (95% CI 20.70-108.86), 0.04 (95% CI 0.02-0.08), and 1524.81 (95% CI 625.29-3718.34), respectively.

In this study, we evaluated the performance of the DL model regarding detection of COVID-19 automatically using chest images to assist with proper diagnosis and prognosis. The findings of our study showed that the DL model achieved high sensitivity and specificity (95% and 96%, respectively) when detecting COVID-19. The pooled SROC value of both COVID-19 and other types of pneumonia was 98%. The performance of the DL model was comparable to that of experienced radiologists, whose clinical experience was at least 10 years, and the model could improve the performance of junior radiologists.

The rate of COVID-19 cases has been mounting day by day; therefore, it is important to quickly and accurately diagnose patients so that we may combat this pandemic. However, screening an increased number of chest images is challenging for the radiologists, and the number of trained radiologists is not sufficient, especially in underdeveloped and developing countries [34]. The recent success of DL applications in imaging analysis of CT scans, as well as x-ray imaging in automatic segmentation and classification in the radiology domain, has encouraged health care providers and researchers to exploit the advancement of deep neural networks in other applications [35]. DL models have been trained to assist radiologists in achieving higher interrater reliability during their years of experience in clinical practice.

Since the start of the COVID-19 pandemic, efforts have been made by AI researchers and AI modelers to help radiologists in the rapid diagnosis of COVID-19 in order to combat the COVID-19 pandemic [33,36]. Developing an accurate, automated AI COVID-19 detection tool is deemed as essential in reducing unnecessary waiting times, shortening screening and examination times, and improving performance. Moreover, such a tool could help to reduce radiologists’ workloads and allow them to respond to emergency situations rapidly and in a cost-effective manner [25]. RT-PCR is considered the gold standard detection method; however, findings of our study showed that chest CT could be used as a reliable and rapid approach for screening of COVID-19. Our findings also showed that the DL model was able to discriminate COVID-19 from other types of pneumonia with high a sensitivity and specificity, which is a challenging task for radiologists [32].

Our study has several strengths. First, this is the first meta-analysis that evaluated the performance of a DL model to classify COVID-19 patients. Second, we considered only peer-reviewed articles to be included in our study because articles that are not peer reviewed might contain bias. Third, we compared the performance of the DL model with that of senior and junior radiologists, which would be helpful for policy makers in considering an automated classification system in real-world clinical settings in order to speed up routine examination.

However, our study also has some limitations that need to be addressed. First, only 16 studies were used to evaluate the performance of the model; inclusion of more studies may have provided more specific findings. Second, some studies included similar data sets, which may have created some bias, but the researchers in those studies had optimized algorithms to improve performance. Third, two different kinds of digital photographs (ie, CT scan and x-ray) were used to develop and evaluate the performance of the DL model in classifying COVID-19; however, the performance of the DL model was almost the same in both cases. Finally, none of the studies included external validation; therefore, model performance could vary if those models were implemented in other clinical settings.

The primary objectives of prediction models are the quick screening of COVID-19 patients and to help physicians make appropriate decisions. Misdiagnosis could have a destructive effect on society, as COVID-19 could spread from infected people to healthy people. Therefore, it is important to select a target population among which this automated tool could serve a clinical need; it is also important to select a representative data set on which the model could be trained, developed, and validated internally and externally. All the studies included in this meta-analysis had a high risk of bias for patient selection and reference standards. Moreover, generalizability was lacking in the newly developed classification models. Models without proper evidence and with a lack of external validation are not appropriate for clinical practice because they might cause more harm than good. Since the number of cases is mounting each day and COVID-19 is spreading to all continents, it is therefore important to develop a model to assist in the quick and efficient screening of patients during the COVID-19 pandemic. This could encourage clinicians and policy makers to prematurely implement prediction models without sufficient documentation and validation. All studies showed promising discrimination in their training, testing, and validation cohorts, but future studies should focus on external validation and comparing their findings to other data sets. Interpretability of DL systems is more important to a health care professional than to an AI expert. Proper interpretation and explanation of algorithms will more likely be acceptable to physicians. More clinical research is needed to determine the tangible benefits for patients in terms of the high performance of the model. High sensitivity and specificity do not necessarily represent clinical efficacy, and the higher value of the AUROC is not always the best metric to exhibit clinical applicability. All papers should follow standard guidelines and they should present positive and negative predictive values in order to be able to make a fair comparison. Although all of the included studies used a significant amount of data to compare model performance to that of the radiologists, they used only retrospective data to train the models, which might result in worse performance in real-world clinical settings, as data complexity is different. Therefore, prospective evaluation is needed in future studies before considering implementation in clinical settings. AI models always consist of potential flaws, including the inapplicability of new data, reliability, and bias. Generalization of the model is important for presenting the real performance because the rate of sensitivity and specificity varied across the studies (0.79 to 1.00 and 0.62 to 1.00, respectively). A higher number of false negatives will make the situation worse and will waste health care resources.

Our study showed that the DL model had immense potential to distinguish COVID-19 patients, with high sensitivity and specificity, from patients with other types of pneumonia and normal patients. DL-based tools could assist radiologists in the fast screening of COVID-19 and in classifying potential high-risk patients, which could have clinical significance for the early management of patients and could optimize medical resources. A higher number of false negatives could have a devastating effect on society; therefore, it is crucial to test the performance of models with other, unknown data sets. Retrospective evaluation and reliable interpretation are warranted to consider the application of AI models in real-world clinical settings.