Endoscopic imaging modalities for diagnosing invasion depth of superficial esophageal squamous cell carcinoma: a systematic review and meta-analysis

We searched the MEDLINE, Cochrane Central, and Ichushi databases from January 1995 to June 2015 using the following search terms: (“esophageal cancer” OR “esophageal tumor” OR “esophageal tumor” OR “esophageal neoplasia” OR “esophageal carcinoma” OR “esophageal mucosal” OR “esophageal lamina propria”) AND (“diagnosis” OR “endosonography” OR “staining and labeling” OR “iodine” OR “magnifying endoscopy OR “chromoendoscopy” OR “NBI” OR “avascular area” OR “endoscopic ultrasound” OR “imaging” OR “pathology” OR “esophagoscopy”) AND (“neoplasm invasiveness” OR “[T1a and EP]” OR “M1” OR “Tis” OR “[T1a and LPM]” OR “M2” OR “T1a” OR “(T1a and MM)” OR “M3” OR “T1b” OR “[pT1a and MM]” OR “T1b” OR “SM” OR “SM1” OR “SM2” OR “SM3” OR “[T1b and SM] OR “vascular involvement” OR invasion OR “infiltration” OR “depth”). Our search was restricted to English- or Japanese-language studies of human subjects. Two reviewers (R.I. and N.M.) independently screened the titles and abstracts of all the articles according to the defined inclusion and exclusion criteria. The final complete report of all selected articles was then retrieved and reviewed by the same two reviewers (R.I. and N.M.). We also manually screened the reference lists of the selected articles for any potential related articles that were not identified by the initial search (Manual searching). Discrepancies were resolved by discussions. The protocol for this meta-analysis was registered in PROSPERO (International Prospective Register of Systematic Reviews; number 42015024462), in accordance with the most recently published guidelines [16].

The study population consisted of patients with esophageal SCC based on endoscopic biopsy and endoscopic examination. The intervention was endoscopic diagnosis (non-ME, ME or EUS) of cancer invasion depth for superficial SCC. The reference standard was histologic diagnosis of cancer invasion depth by ER, or from surgically resected specimens. Acceptable study designs were retrospective or prospective studies with sufficient data to allow reconstruction of a diagnostic 2 × 2 table (true positive, false positive, true negative, and false negative). We excluded case reports, review articles, and studies in which the total number of patients or lesions was <10. We also excluded studies that did not provide any predefined criteria to diagnose invasion depth and studies with imaging modalities that are not used in daily practice.

Histologic diagnosis of cancer invasion depth was divided into six categories, based on the findings: EP (cancer limited to the epithelium); LPM (cancer invading into the lamina propria); MM (cancer invading into the muscularis mucosa); SM1 (cancer invading 0.2 mm below the lower border of the muscularis mucosa in endoscopically resected specimens and cancer invading the upper third of the submucosal layer in surgically resected specimens); SM2 (cancer invading >0.2 mm into the submucosa in endoscopically resected specimens and cancer invading the middle third of the submucosal layer in surgically resected specimens); SM3 (cancer invading the lower third of the submucosal layer in surgically resected specimens) [17].

Endoscopic diagnosis of cancer invasion depth was divided into three categories: EP/LPM, MM/SM1, and ≥ SM2, because these categories correspond well with the risk of metastasis [10] and indication of ER. Moreover, most diagnostic criteria for cancer invasion depth of esophageal SCC were developed to differentiate these three categories, and there are currently no popular non-ME or ME criteria for differentiating between mucosal and submucosal cancers.

Two independent reviewers (R.I. and N.M) extracted the following data from the selected studies and added them to standardized data forms: design; country; year of publication; setting; sample size; reference standard; operating frequencies of endoscope and/or probe; number of endoscopic imaging modalities used; and numbers of true-positive, true-negative, false-positive and false-negative values.

Study quality and potential bias were assessed according to the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool [18], which included four key domains: patient selection, index test, reference standard, and flow timing. Each domain was assessed for risk of bias, and the first three domains were also assessed regarding applicability. Quality assessment of the studies was performed independently by R.I. and N.M, and any disagreement was resolved by discussion.

We constructed 2 × 2 tables for EP/LPM and ≥ MM, and for EP-SM1 and ≥ SM2 for each study, based on comparisons between the endoscopic diagnosis and final histologic diagnosis by ER or esophagectomy. The true-positive, false-positive, true-negative, and false-negative values were then calculated based on the 2 × 2 tables. A summary receiver operating characteristic curve (SROC) was constructed [19]. A SROC is similar to a standard ROC, except that the SROC data are obtained from the sensitivity and specificity values in the individual studies in the meta-analysis. The area under the curve (AUC) of a SROC is an indicator of the performance of a diagnostic modality [19]. A preferred test has an AUC close to 1, and a poor test has an AUC close to 0.5 [20]. The Q* index is the point where the sensitivity and specificity are equal, which is the point closest to the ideal top-left corner of the SROC space [19].

The pooled sensitivity, specificity, PLR, NLR, and diagnostic odds ratio were estimated using a fixed-effect model (Mantel–Haenszel method). Forest plots were used to show the effect size of each study. Heterogeneity was assessed using Cochran’s Q test and the I² measure of inconsistency [21‐23]. The Cochran Q test detects heterogeneity by testing the null hypothesis that all studies in a meta-analysis have the same underlying magnitude of effect. Because this test is underpowered to detect moderate degrees of heterogeneity, a P value of <0.10 was considered suggestive of significant heterogeneity [24]. The I² index describes the percentage of total variation among studies attributed to heterogeneity rather than chance. A value of 0% indicates no observed heterogeneity, and larger values show increasing heterogeneity. Higgins et al. [21] suggested that I² indexes of 25%, 50%, and 75% represented low, moderate, and high heterogeneity, respectively. For all statistical methods, except for Cochran’s Q test, P <0.05 was regarded as significant. Data were analyzed using Meta-Disc (version 1.4) and Review Manager.

A total of 359 articles were initially identified using the search strategy, and 18 additional records were identified through manual searching of references (Fig. 1). Among all the studies, 300 were excluded after preliminary review of the titles and abstracts, leaving 77 articles for detailed evaluation. Of these, 63 articles failed to meet the criteria and 14 studies were finally selected for this meta-analysis [25‐38]. Only two of these were prospectively designed studies [27, 36]. All of 14 were Japanese studies and 11 of them were written in Japanese. A total of 359, 1613 and 357 patients received non-ME, ME and EUS, respectively. Details of the studies are described in Table 1 [12, 13, 39].

Summary ROC curves showed that ME and EUS were positioned in the upper right corner of the ROC space compared with non-ME (Fig. 2a, b). The AUC was used to summarize the overall diagnostic accuracy of each modality. Non-ME, ME, and EUS had AUC values of 0.934, 0.946, and 0.975, respectively, for differentiating between EP/LPM and ≥ MM, while ME and EUS had AUC values of 0.999 and 0.966, respectively, for differentiating between EP-SM1 and ≥ SM2.

The Forest plots of sensitivity, specificity, PLR, and NLR for each modality for differentiating between EP/LPM and ≥ MM, and between EP-SM1 and ≥ SM2 are shown in Fig. 3a–d and e–h, respectively. Point estimates with 95% confidence intervals (CIs) were plotted for each group (Fig. 3 a-h). ME had significantly higher sensitivities for diagnosing EP/LPM (0.96 [95%CI: 0.91–0.96]) and EP-SM1 cancers (1.00 [95%CI: 0.99–1.00]) compared with non-ME and EUS. ME also had very low NLR for diagnosing EP/LPM (0.08 [95%CI: 0.03–0.25]) and EP-SM1 cancers (0.01 [95%CI: 0.00–0.02]). EUS showed significantly higher specificities for the diagnosis of EP/LPM (0.97 [95%CI: 0.93–0.99]) and EP-SM1 cancers (0.94 [95%CI: 0.98–0.88]) compared with non-ME and ME. EUS also had a very high PLR for diagnosing EP/LPM (17.63 [95%CI: 6.71–46.34]) and EP-SM1 cancers (11.60 [95%CI: 5.44–24.74]).

The qualities of the included studies evaluated according to the QUADAS-2 criteria are shown in Fig. 4. Half the studies showed risk of bias regarding “Patient selection” and “Flow and timing”, mainly as a result of unclear descriptions of the patient-selection process and analysis methods. The Cochran Q test identified heterogeneities for differentiating between EP/LPM and ≥ MM by non-ME (P = 0.076 for sensitivity and P = 0.002 for specificity) and ME (P = 0.002 for sensitivity and P < 0.001 for specificity), between EP-SM1 and ≥ SM2 by ME (P < 0.001 for specificity). The I² index identified moderate to high heterogeneities for differentiating between EP/LPM and ≥ MM by non-ME (61.2% for sensitivity and 83.5% for specificity) and ME (91.3% for sensitivity and 87.7% for specificity), between EP-SM1 and ≥ SM2 by ME (91.2% for specificity). Sensitivity analysis was not performed because of the limited number of studies of each modality. However, heterogeneity for differentiating between EP/LPM and ≥ MM by non-ME was resolved by excluding one study [27], and heterogeneity for differentiating between EP-SM1 and ≥ SM2 by ME was resolved by excluding another study [31].