This study was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Review and Meta-Analyses) of DTA Studies [15]. The study protocol was registered at the International Prospective Register of Systematic Reviews database before the initiation of the systematic review (CRD42021232189). Approval from the institutional review board of the Chuncheon Sacred Heart Hospital was waived.

Two authors (CSB and JJL) independently performed a core database search of MEDLINE, PubMed, Embase, and Cochrane Library using common search formulas, from inception to January 2020. Duplicate articles were excluded from the analyses. The titles and abstracts of all identified articles were reviewed, and irrelevant articles were excluded. Full-text reviews were subsequently performed to determine whether the pre-established inclusion criteria were satisfied in the identified literature. References were also reviewed to identify any additional relevant articles. Any disagreements in the results of the search process between the 2 authors were resolved by discussion or consultation with a third author (GHB). The search formulas used to identify the relevant articles are presented in Textbox 1.

The literature included in this systematic review should meet the following inclusion criteria: designed to evaluate the diagnostic performance of CAD models in the prediction of histology of DCPs based on endoscopic images; presentation of the diagnostic performance of CAD models, including sensitivity, specificity, likelihood ratios, predictive values, or accuracy, which enabled the estimation of true-positive (TP), false-positive (FP), false-negative (FN), and true-negative (TN) values for the histologic diagnosis of DCPs based on endoscopic images; and studies written in English. The exclusion criteria were as follows: narrative review articles; studies with incomplete data; systematic reviews or meta-analyses; and comments, proceedings, or study protocols. Articles meeting at least one of the exclusion criteria were excluded from this systematic review.

Two authors (CSB and JJL) assessed the methodological quality of the final included articles using the second version of the Quality Assessment of Diagnostic Accuracy Studies. This tool comprises four domains: patient selection, index test, reference standard, and flow and timing, and the first three domains have an applicability assessment. The 2 authors assessed each part as having a high, low, or unclear risk of bias [16].

Two authors (CSB and JJL) independently extracted the data from each included study and cross-checked the extracted data. If the data were unclear, the corresponding author of the study was contacted by email to obtain insight into the original data set. A descriptive synthesis was performed using a systematic review process, and DTA meta-analysis was conducted if the included studies were sufficiently homogenous.

The primary outcomes were the TP, FP, FN, and TN values in each study. For the CAD of the histology of DCPs using endoscopic images, the primary outcomes were defined as follows: TP referred to the number of subjects with a positive finding by a CAD model and have adenomas or neoplasms as evidenced by endoscopic images; FP referred to the number of subjects with a positive finding by a CAD model and do not have adenomas or neoplasms based on endoscopic images; FN referred to the number of subjects with a negative finding by a CAD model and have adenomas or neoplasms as evidenced by endoscopic images; and TN referred to the number of subjects with a negative finding on a CAD model and do not have adenomas or neoplasms based on the endoscopic images. With these definitions, the TP, FP, FN, and TN values were calculated for each included study. If the included studies presented comparative diagnostic performance of endoscopists versus CAD models, the TP, FP, FN, and TN values of endoscopists in each study were also extracted.

For additional analyses, such as subgroup analysis or meta-regression, the authors extracted the following variables from each included study: publication year, geographic origin of the data (ie, Western vs Asian), type of endoscopic images, type of CAD models, location of the DCPs (ie, any colon vs rectosigmoid colon), number of total images included, and type of test data sets (internal test vs external test).

The bivariate method [17] and hierarchical summary receiver operating characteristic (HSROC) method [18] were applied for the DTA meta-analysis. A forest plot of the sensitivity and specificity and a summary receiver operating characteristic (SROC) curve were generated using the bivariate method [17] and HSROC [18] method, respectively. The level of heterogeneity across the included articles was determined by the correlation coefficient between logit-transformed sensitivity and specificity by the bivariate method and the asymmetry parameter β, where β=0 corresponds to a symmetric receiver operating characteristic curve, in which the diagnostic odds ratio (DOR) does not vary along the curve according to the HSROC method. A positive correlation coefficient and a β with a significant probability (P<.05) indicated heterogeneity between the studies [18,19]. A visual examination of the SROC curve was also performed to identify heterogeneity. Subgroup analysis by univariate meta-regression using the modifiers identified during the systematic review was also performed to identify the reasons for heterogeneity. The pooled NPV by integrating conditional prevalence with respect to a previous distribution (considering the heterogeneity in prevalence) was calculated using a probability-modifying plot.

STATA software version 15.1, including the METANDI and MIDAS packages, was used for the DTA meta-analysis. The METANDI and MIDAS packages require the inclusion of a minimum of four studies for DTA meta-analysis. Therefore, if less than four studies were included in the subgroup analysis, the Moses-Shapiro-Littenberg method [20], as implemented in Meta-DiSc 1.4 (XI Cochrane Colloquium), was used. Publication bias was evaluated using the Deek funnel plot asymmetry test.

A total of 277 articles were identified following a literature search of the three core databases. Two additional studies were identified by manual screening of the bibliographies. After excluding 111 duplicate studies, 57 additional articles were excluded after reviewing the titles and abstracts. Full-text versions of the remaining 111 articles were obtained and thoroughly reviewed based on the aforementioned inclusion and exclusion criteria. Among these, 98 articles were excluded from the final enrollment for the following reasons: 73 (74%) for incomplete data, 11 (11%) for narrative review, 8 (8%) for study protocol, and 6 (6%) for systematic review or meta-analysis. Finally, 13 studies [21-33] were included in the systematic review. A flowchart of the selection process is presented in Figure 1.

The identified studies established and explored the diagnostic performance of CAD models for the classification of adenomatous or neoplastic versus nonadenomatous or nonneoplastic polyps. Among the 13 studies for the CAD of DCPs, 6564 images were identified (3207 cases vs 3357 controls).

Seven studies [24,26,27,29-32] used endoscopic images from Asian populations, and six studies [21-23,25,28,33] used endoscopic images from Western populations. All included studies adopted the definition of DCPs as size <5 mm. However, Shahidi et al [23] adopted a stricter definition for DCPs with a size <3 mm. With regard to the type of CAD model, a deep neural network or convolutional neural network was used in six studies [21-25,30], a support vector machine in six studies [27-29,31-33], and a software-based automatic color intensity analysis in one study [26]. White-light imaging is currently the standard method for inspecting endoscopic lesions. However, one study [28] used white-light imaging to establish a CAD model, and most of the included studies used image-enhanced endoscopic images, such as narrow-band imaging [21-26,30,32,33] or autofluorescence imaging [26] with or without magnification for the detailed characterization of the morphology of DCPs. Three studies [27,29,31] have used images of endocytoscopy, which is a specialized endoscopy that allows the analysis of mucosal structures at the cellular level [34]. With regard to the location of DCPs, most studies [21,23-25,27,28,30-33] did not consider the location of DCPs. However, three studies [22,26,29] separately collected DCPs from the rectosigmoid colon and evaluated the diagnostic performance of the CAD model for these polyps. Most of the included studies [21,23-26,28-33], except for two studies [22,27], have evaluated diagnostic performance using an internal test data set. A study by Zachariah et al [22] presented both external and internal test performance, and a study by Kudo et al [27] presented an external test format of the CAD model previously established by Mori et al in 2016 [31] and 2018 [29,31]. Five studies [25,27,30,31,33] have presented the comparative diagnostic performance of endoscopists versus CAD models for the prediction of histology in DCPs. Detailed clinical features of the included studies are presented in Table 1.

The quality of the baseline image data is important because the CAD model is established using the learning features of the baseline training data. Theoretically, the images included in each study should reflect real-world conditions, as the CAD model was established for use in clinical practice. However, as some lesions are rare or abnormal, data imbalance is the main barrier to the learning of CAD models. Most of the included studies in the systematic review attempted to mitigate this pitfall by adopting specific inclusion and exclusion criteria for the enrollment of endoscopic images. However, four studies [21,23,25,28] did not include a detailed description of the image enrollment standard. Therefore, these studies were rated as unclear risk in the patient selection domain (Figures 2 and 3). This binary classification of low risk and unclear risk in the patient selection domain was adopted as a modifier in the subgroup or meta-regression analysis.

Among the 13 studies [21-33] for the meta-analysis of CAD of DCPs, the area under the curve (AUC), sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, and DOR of CAD models for the diagnosis of DCPs were 0.96 (95% CI 0.93-0.97), 0.93 (95% CI 0.91-0.95), 0.87 (95% CI 0.76-0.93), 7.1 (95% CI 3.8-13.3), 0.08 (95% CI 0.06-0.11), and 87 (95% CI 38-201), respectively (Figure 4; Table 2). The SROC curve is shown in Figure 5. To investigate the clinical utility of the CAD models, Fagan nomogram was generated. Positive findings indicated that adenomas or neoplasms were detected by the CAD models. Negative findings indicated that nonadenomas or nonneoplasms were detected by the CAD models. After assuming a 49% prevalence of adenomas or neoplasms (this value was calculated from the values in Table 1; ie, the total number of cases/controls, ie, 3207/3357, 95.53%), the Fagan nomogram shows that the posterior probability of adenomas or neoplasms was 87% if the finding of the CAD model was positive, and the posterior probability of adenoma was 7% if the finding of the CAD model was negative (Figure 6).

Five studies [25,27,30,31,33] compared the performance of CAD models and endoscopists. Among these, four studies [27,30,31,33] have presented comparative performance between CAD models and endoscopists according to the expertise of the endoscopists (expert endoscopists vs CAD models or novice endoscopists vs CAD models). The pooled AUC, sensitivity, specificity, and DOR of CAD models for the diagnosis of DCPs were 0.96 (95% CI 0.94-0.98), 0.96 (95% CI 0.95-0.97), 0.91 (95% CI 0.84-0.95), and 231 (95% CI 113-473), respectively. For the expert endoscopists, the pooled AUC, sensitivity, specificity, and DOR were 0.97 (95% CI 0.95-0.98), 0.93 (95% CI 0.89-0.96), 0.89 (95% CI 0.80-0.95), and 116 (95% CI 80-168), respectively. For the novice endoscopists, the pooled AUC, sensitivity, specificity, and DOR were 0.85 (95% CI 0.82-0.88), 0.78 (95% CI 0.68-0.86), 0.78 (95% CI 0.68-0.86), and 13 (95% CI 6-30), respectively. The forest plot of AUCs is illustrated in Figure 7, and no significant difference was found between CAD models and expert endoscopists; however, novice endoscopists showed lower pooled AUC for the histologic diagnosis of DCPs than those for CAD models or expert endoscopists.

With regard to the NPV of CAD models for the diagnosis of adenomatous polyps in the rectosigmoid colon, the pretest prevalence of adenomatous polyp in the rectosigmoid colon was 13.2% (95% CI 10.2%-16.5%) in a recent meta-analysis [35]. For the assumption of this prevalence, the NPV of CAD models was 0.99 (95% CI 0.87-0.99; Figure 8). In this meta-analysis, the prevalence of adenomatous polyp in the rectosigmoid colon was 30% (95% CI 27%-32%) based on 29.53% (352/1192) of polyps in the rectosigmoid colon. For the assumption of this prevalence, the NPV of CAD models was 0.97 (95% CI 0.87-0.99; Figure 9). If we adopt a simple follow-up equation for the NPV of CAD models for the diagnosis of adenomatous polyps in the rectosigmoid colon using pooled sensitivity and specificity, the NPV of CAD models was 0.96 (95% CI 0.95-0.97). The follow-up equation is as follows:

NPV = (specificity × [1−prevalence])/(specificity × [1−prevalence] + prevalence × [1−sensitivity])

First, the authors observed a positive correlation coefficient between the logit-transformed sensitivity and specificity (r=0.22) and an asymmetric β parameter, with a significant P value (P=.004), implying that heterogeneity exists among the studies. Second, a coupled forest plot of sensitivity and specificity was obtained (Figure 4). Compared with the enrolled studies, the study by Shahidi et al [23] showed lower specificity. This study was found to have an unclear risk of bias in methodology quality assessment. Therefore, subgroup analysis was carried out according to the methodological quality, and a negative correlation coefficient was found between logit-transformed sensitivity and specificity (r=−0.13) and an asymmetric β parameter, with a nonsignificant P value (P=.63) in high-quality studies, indicating an absence of heterogeneity among the studies. Third, the shape of the SROC curve for CAD of DCPs using endoscopic images was symmetric (Figure 5). Fourth, meta-regression using modifiers identified in the systematic review was conducted, and no source of heterogeneity could be identified (published year, P=.34; nationality of the data sets, P=.29; type of CAD models, P=.38; type of endoscopic image, P=.23; location of the DCPs, P=.90; type of test data sets, P=.66; total number of images, P=.66; and methodological quality, P=.10; Figure 10). Finally, a subgroup analysis based on the potential modifiers was performed, and the pooled AUC of studies with high methodological quality was higher than that of studies with lower methodological quality. Except for this variable (methodological quality), no significant changes in diagnostic performance were found according to the modifiers (Table 2).

Deek funnel plot of studies for the CAD of DCPs exhibited a symmetrical shape with respect to the regression line (Figure 11), and the asymmetry test showed no evidence of publication bias (P=.65).

This study presented evidence that CAD models showed high performance values for the histologic diagnosis of DCPs and practical values in the Fagan nomogram, indicating the potential to use these models in clinical practice. This performance was comparable with that of expert endoscopists and higher than that of novice endoscopists. Although the main analysis found heterogeneity among the included studies, the subgroup analysis demonstrated that methodological quality was the reason for the heterogeneity. Thorough meta-regression or subgroup analyses did not reveal any additional reasons for heterogeneity.

Most polyps detected during colonoscopy are diminutive, and considering the low potential of malignancy, resecting all DCPs is not cost-effective [4,7-9]. However, many DCPs are still being resected and sent for histologic evaluation to determine the surveillance interval for CRC screening [10]. CAD models without pathologic diagnosis could lead to cost savings by changing surveillance interval recommendations, and the DTA meta-analysis in our study revealed that the NPV of CAD models for the diagnosis of adenomatous polyps in the rectosigmoid colon was over 90% to adopt the diagnose and leave strategy based on the optical biopsy satisfying PIVI performance thresholds for in situ endoscopic histology prediction of DCPs.

Despite the technical challenges of CAD models analyzing a smaller surface area, previous meta-analyses [36-40] have demonstrated that CAD models can increase the adenoma or polyp detection rate, especially for those with small size. With regard to the histologic prediction of DCPs, a previous meta-analysis revealed that endoscopists with only high confidence showed an NPV of approximately 90% using CAD models of digital chromoendoscopy, implicating the potential for adopting the diagnosis and leave strategy [41]. Another meta-analysis showed an NPV of 0.95 (95% CI 0.88-0.98) for the CAD of DCPs in nonmagnifying narrow-band imaging [42]. However, the location of the DCPs was not considered, and many studies were omitted from the search process.

An additional finding of this DTA meta-analysis is the robustness of the diagnostic performance of CAD models. The performance values were consistent regardless of the modifiers, except for the methodological quality (Table 2). This was consistent regardless of the nationality of the patients, location of DCPs, total number of included polyps, and type of CAD models or endoscopic images. However, the diagnostic performance of studies with high methodological quality showed higher AUCs than that of studies with low methodological quality, and no evidence of heterogeneity was detected among studies in the subgroup with high methodological quality. Although pooled AUCs in the subgroup of external test datasets could not be measured because only two studies were included in this subgroup, the remaining performance values were comparable with the subgroup of internal test data sets.

Despite the robust evidence in the DTA meta-analysis stated earlier, several inevitable limitations were identified. First, only two or three studies were included in the subgroup analyses of external test data sets, rectosigmoid DCPs, and endocytoscopic images. Bivariate and HSROC methods are advanced statistical techniques that have overcome the limitations of the Moses-Shapiro-Littenberg method (which does not consider any heterogeneity between studies) [43,44]. However, the Moses-Shapiro-Littenberg method is only possible for a subgroup with fewer than four studies. With accumulating evidence on this topic, this statistical pitfall could be overcome. Second, the number of studies was insufficient to enable a comparison of the relative diagnostic performance of endoscopists and CAD models. Considering the real clinical adaptation of endoscopists with CAD models rather than endoscopists versus CAD models, it is no longer necessary to compare the diagnostic capabilities of doctors and CAD models [43]. Owing to the unique characteristics of patients in each institution, CAD models developed from a single institution usually have limitations for widespread implementation, indicating the importance of the external test. However, only two studies conducted external tests to verify CAD model performance. Additional studies focusing on external validation-oriented performance or suggesting a clinical application benefit for future perspectives in established CAD models are expected.

In conclusion, CAD models showed potential for the optical histological diagnosis of DCPs using endoscopic images.