Introduction
Gastric cancer (GC) is one of the major cancer burdens globally [1]. Despite the fact that the trend in death rates is decreasing, GC still has a poor prognosis and few therapeutic options are efficacious, particularly in advanced stages of the disease. It is now well established that GC pathogenesis is a multifactorial process, in which Helicobacter pylori, environmental and host factors may play significant roles [2]. It is a multistep process that includes the sequential development of chronic gastritis, followed by gastric mucosal atrophy, gastric intestinal metaplasia (GIM), dysplasia, and ultimately adenocarcinoma [2-4]. GIM is generally considered as a pre-cancerous lesion in the gastric mucosa. However, when white light endoscopy (WLE) is used, the detection of GIM requires biopsies for proper histology assessment. Fortunately, the detection of GIM is possible using improved endoscopic techniques [5].
Narrow band imaging (NBI) involves optical image enhancement that uses 2 short wavelength light beams, 415 nm (blue) and 540 nm (green) [6]. It is an endoscopic technique that can provide enhanced visualization of the micro-surface structure and microvascular architecture of subjacent microvascular patterns [7]. Consequently, studies have examined the association between the appearance of the mucosa observed with NBI and pathology [8-10]. In particular, the combination of NBI with magnification is of importance, since it facilitates the detection of the characteristic light blue crest, thus increasing the possibility of diagnosing GIM [11].
So far, there have been 2 meta-analyses addressing the role of NBI in diagnosing GIM. One was published some years ago (2014) [12] and the other was a pairwise meta-analysis comparing NBI to WLE, expressing the comparison results as odds ratios and not addressing the diagnostic accuracy [13]. Therefore, the aim of this systematic review and diagnostic accuracy meta-analysis was to evaluate the current diagnostic accuracy of NBI in detecting GIM.
Materials and methods
Selection criteria
We strictly followed the PRISMA guidelines for systematic reviews [14] and potentially eligible studies for the meta-analysis were selected by applying the following inclusion criteria: (a) they were written in the English language; and (b) they contained data that were appropriate for the construction of 2-by-2 contingency tables for calculations of NBI diagnostic accuracy parameters. When the same data were reported in 2 studies the more informative study was selected. QUADAS-2 evaluation was used for quality estimation of the eligible studies [15].
Identification of studies and data extraction
We searched PubMed/MEDLINE and EMBASE databases for suitable studies comparing the endoscopic diagnosis of GIM with histology (Fig. 1). We used the following search terms: “NBI”[All Fields] AND (“intestinalization”[All Fields] OR “intestinalized”[All Fields] OR “intestinally”[All Fields] OR “intestinals”[All Fields] OR “intestine s”[All Fields] OR “intestines”[MeSH Terms] OR “intestines”[All Fields] OR “intestinal”[All Fields] OR “intestine”[All Fields]) AND (“metaplasia”[MeSH Terms] OR “metaplasia”[All Fields] OR “metaplasias”[All Fields]) AND (“sensitivity and specificity”[MeSH Terms] OR (“sensitivity”[All Fields] AND “specificity”[All Fields]) OR “sensitivity and specificity”[All Fields] OR (“sensitivity”[All Fields] AND “specificity”[All Fields]) OR “sensitivity specificity”[All Fields]) AND (“diagnosis”[MeSH Terms] OR “diagnosis”[All Fields] OR “diagnostic”[All Fields] OR “diagnostical”[All Fields] OR “diagnostically”[All Fields] OR “diagnostics”[All Fields]) AND (“accuracies”[All Fields] OR “accuracy”[All Fields]). We searched until the end of September 2022. All selected studies were further screened, searching for more appropriate studies. Data from each study were extracted independently by the 2 authors and any disagreement was settled by discussion.
Figure 1 Flow chart describing the identification, screening, eligibility and inclusion of the studies in this meta-analysis
Statistical analysis
For statistical analysis we followed a methodology described previously [16]. In brief, pooled data concerning all NBI diagnostic accuracy parameters, i.e., sensitivity, specificity and in addition positive and negative likelihood ratios (LR), diagnostic odds ratios (DOR) and AUCs, with 95% confidence intervals (CIs), were derived by computing data obtained from the individual studies and constructing 2-by-2 contingency tables. For calculations we used the fixed-effects model (Mantel and Haenszel method) [17]. When significant heterogeneity was present, we used the random-effects model (DerSimonian and Laird method) [18]. Forest plots were constructed for visual display of individual and pooled data. In addition, the results of the individual studies were displayed in a receiver operating characteristic (ROC) graph, illustrating the distribution of sensitivities and specificities. Additionally, a weighted symmetric summary ROC (sROC) curve was derived, with calculation of the relevant areas under the curve (AUC). The AUC in accurate tests approaches 1 and in poor tests is close to 0.5 [19-21]. The heterogeneity between studies was estimated using the Cochran Q-test, and the inconsistency index I squared (I2) was used to quantify the degree of heterogeneity [22]. When the Q-test provided a P-value of less than 0.1 and the I2 was more than 50 [23], then heterogeneity was considered to be present, and further sensitivity-subgroup analyses were performed in addition to the random-effects model. The possibility of publication bias was examined using Deeks’ funnel plot, with a superimposed regression line [24]. All analyses were performed with Stata software (version 13.0, College Station, TX) using the MIDAS command and Meta-DiSc software [21].
Results
Study characteristics and descriptive assessment
Fig. 1 shows the flow chart describing the process of study selection. Of 1314 titles initially generated by the literature searches, 11 prospective cohort studies in adult patients [25-35] were found eligible for meta-analysis. These studies included a total of 1672 patients whose NBI endoscopy GIM findings were compared to the gold standard, i.e., histology. The main characteristics of the 11 included studies are shown in Table 1. The quality assessment of the included studies according to the QUADAS-2 evaluation is shown in Supplementary Fig. 1.
Table 1 Main characteristics of the studies included in the meta-analysis
NBI diagnostic performance
In view of the significant heterogeneity found for both the sensitivity and the specificity, with respective heterogeneity values of Q-test=49.63, d.f.=10, P<0.001, I2=79.85% and Q-test=39.87, d.f.=10, P<0.001, I2=74.92%, the random-effects analysis was used throughout the calculations. The pooled data showed that NBI had a sensitivity of 80% (95%CI 69-87%) and a specificity of 93% (95%CI 85-97%), with respective forest plots shown in Fig. 2.
Figure 2 Forest plot showing sensitivities (A) and specificities (B) with corresponding heterogeneity statistics
CI, confidence interval
The corresponding ROC plot with sROC is displayed in Fig. 3A, showing an AUC of 0.93 (0.91-0.95). The possibility of publication bias was explored using Deek’s funnel plot asymmetry test with a superimposed regression line, as shown in Fig. 3B. There was no significant evidence of publication bias (P=0.88 for the slope coefficient). In addition, the bivariate boxplot is shown in Fig. 3C, where most studies are clustered within the median distribution with some outliers, indirectly suggesting the magnitude of the heterogeneity. The relevant LR scattergram is shown in Fig. 3D and provides the summary point of LRs obtained as functions of mean sensitivity and specificity. In exploring the reasons for the significant heterogeneity found among studies, further analyses (sensitivity analyses) were conducted. These are shown in Fig. 4A-D, depicting the residual-based goodness-of-fit, the bivariate normality, the influence analysis and the outlier detection, respectively. These analyses identified 3 outlier studies that contributed to the significant heterogeneity found. Fig. 5A shows the respective Fagan’s nomogram providing 73% post-test probability of GIM after an NBI-positive result and only a 5% post-test probability after an NBI-negative result. The relevant probability modifying plot is shown in Fig. 5B, with a positive LR 10.63 (95%CI 5.07-22.32) and negative LR 0.22 (95%CI 0.14-0.34). These results give a 90% (95%CI 86-95%) positive predictive value (PPV) and an 80% (95%CI 76-84%) negative predictive value (NPV).
Figure 3 (A) Weighted symmetric summary receiver operating curve (sROC), with 95% confidence intervals and prediction regions around mean operating sensitivity and specificity point. (B) Deeks’ funnel plot, with superimposed regression line. No evidence of publication bias. (C) Bivariate box plot, with most studies clustering within the median distribution and some outliers indirectly suggesting the existence of heterogeneity. (D) Likelihood ratio scattergram
Figure 4 Sensitivity analyses, with graphical depiction of: (A) residual-based goodness-of-fit, (B) bivariate normality, (C) influence analyses, (D) outlier detection analyses
Figure 5 (A) Fagan’s nomogram showing pre-test and post-test probabilities with relevant likelihood ratios. (B) Probability modifying plot
Subgroup analysis
Given the significant heterogeneity found, in addition to sensitivity analyses, subgroup analysis was performed according to the NBI modality used, i.e., with and without magnification. NBI with magnification was used in 6 studies [25,28-30,34,35], whereas 5 studies [26,27,31-33] used NBI without magnification. The pooled data (random-effects analysis) showed that NBI with magnification performed better than NBI without magnification (Table 2). For NBI with magnification, sensitivity (95%CI), specificity, LR positive, LR negative, DOR, ROC area, PPV and NPV were 0.85 (95%CI 0.74-0.92), 0.93 (95%CI 0.86-0.97), 11.97 (95%CI 5.69-25.21), 0.16 (95%CI 0.09-0.30), 75 (95%CI 24-239), 0.95 (95%CI 0.93-0.97), 0.91 (95%CI 0.87-0.95), 0.85 (95%CI 0.81-0.88), respectively. For NBI without magnification the respective values were 0.70 (95%CI 0.55-0.82), 0.94 (95%CI 0.66-0.99), 12.24 (95%CI 1.74-86.11), 0.31 (95%CI 0.20-0.49), 39 (95%CI 5-277), 0.84 (95%CI 0.81-0.87), 0.91 (95%CI 0.84-0.98), 0.74 (95%CI 0.69-0.80).
Table 2 NBI sensitivity, specificity, LR positive, LR negative, DOR, ROC area, PPV, NPV in all included studies and the 2 subgroups, i.e., with and without magnification
Discussion
GC is the third leading cause of cancer deaths worldwide, with 1,033,701 new cases diagnosed globally in 2018 [1]. GIM has been associated with an increased risk for GC and may represent the histological step before the development of dysplasia [36]. Furthermore, it has been considered as a specific marker to identify patients who might benefit from surveillance. The ability to identify precursor lesions on gastric biopsies has led to interest in developing screening and surveillance strategies for the early detection and prevention of GC. Various endoscopic techniques have been developed for diagnosing mucosal lesions, but not all of them are widely available. In standard or high definition WLE, the shorter wavelength can only penetrate superficial mucosa, which limits the ability to evaluate the mucosa in detail. Therefore, WLE does not provide mucosal or microvascular details and, as GIM does not manifest easily detectable lesions, it can be missed when WLE is used alone. Hence, endoscopists usually rely only on random biopsies for the detection of GIM.
The NBI endoscopic technique is based on a modification of the spectral characteristics of the optical filter in the light source, leading to improved visibility of mucosal structures. In addition, NBI does not require tedious patient preparation, as in other modalities such as chromoendoscopy, and it can be used easily by switching on a filter on the endoscope during endoscopy. Consequently, NBI has made possible the detection of mucosal changes that precede malignant changes, such as GIM [6]. In particular, the presence of a blue-white line on the crests of the epithelial surface of the gastric mucosa (the so called “light blue crest” sign) on NBI with magnification has high sensitivity and specificity for GIM [11,25,27-30,35]. Therefore, detection of light blue crest lesions on NBI can enhance the GIM detection rate.
This study addressed the accuracy of NBI in diagnosing GIM, taking into account all the available published evidence regarding the diagnostic performance of NBI in comparison to histology, and the results showed that good diagnostic accuracy was achieved. Thus, the pooled sensitivity was 80% (95%CI 69-87), the pooled specificity was 93% (95%CI 85-97) and the relevant AUC was 0.93 (95%CI 0.91-0.95) in detecting GIM. These data indicate a good diagnostic accuracy for NBI in diagnosing GIM. When we split the whole group of included studies according to the NBI modality used, i.e., with and without magnification, we found that NBI with magnification performed better than NBA in the whole group and the group without NBI magnification, with values for sensitivity, specificity and AUC 0.85 (95%CI 0.74-0.92), 0.93 (95%CI 0.86-0.97), and 0.95 (95%CI 0.93-0.97), respectively.
The ability to identify precursor lesions, such as GIM, on gastric biopsies has led to interest in developing screening and surveillance strategies for the early detection and prevention of GC. Histology is currently considered to be the gold standard diagnostic tool in diagnosing GIM. However, GIM lesions can be missed by random biopsy sampling, since their distribution in various parts of the stomach can be patchy. In addition, histology can carry a financial burden. Consequently, identification of GIM lesions, especially by NBI endoscopy, can be of great diagnostic importance in daily clinical practice, since this endoscopic technique is spreading and endoscopists’ expertise in detecting GIM is improving, which could overcome the need for histology.
The lack of publication bias strengthens the results of this meta-analysis. However, the study has some limitations. Thus, significant heterogeneity was observed among studies. This could be due to a number of factors related to the NBI examination itself, since it is operator dependent and can therefore vary between operators and centers. Moreover, the NBI modality, i.e., NBI with and without magnification, could be of importance. Indeed, when we split the whole group according to the NBI modality used, we found that NBI with magnification performed better than NBI without magnification. Furthermore, the heterogeneity found may be related to the quality of selected studies. Consequently, more data from high quality multicenter studies are necessary to further evaluate whether NBI stands as an accurate endoscopic diagnostic tool in detecting GIM and thus overcoming histology. Finally, an additional limitation might be related to the fact that we included only studies published in English; therefore, language bias may exist.
In conclusion, this meta-analysis showed that NBI is an accurate and useful tool to diagnose GIM. NBI with magnification performed better than NBI without magnification and therefore it should be preferred. However, as only a few studies were available, more high-quality trials are warranted to further evaluate the actual diagnostic accuracy and clinical relevance of NBI for diagnosing GIM.
What is already known:
-
Gastric intestinal metaplasia (GIM) is generally considered as a precancerous lesion in the gastric mucosa
-
Using white light endoscopy, GIM detection requires biopsies for a proper histological assessment
-
Narrow band imaging (NBI) enhances visualization of microvascular architecture and micro-surface structure between the epithelial surface and subjacent vascular pattern with increased likelihood of detecting GIM
What the new findings are:
-
NBI is an accurate and useful endoscopic diagnostic tool in detecting GIM thereby avoiding the need for histology
-
NBI with magnification performed better than NBI without magnification and it should be preferred