We performed this systematic review and meta-analysis according to the Synthesizing Evidence from Diagnostic Accuracy TEsts (SEDATE) guidelines [12]. All methods for inclusion/exclusion criteria, data extraction and quality assessment were specified in advance. The protocol did not require registration.

Studies published between 1989 and January 2017 were identified by 3 of the authors (JLA, BG, BN) using 3 electronic databases (PubMed/MEDLINE, Cochrane, and Web of Science), to identify potentially eligible studies. For ongoing clinical trials, we searched in websites such as www.ClinicalTrials.gov and www.who.int/trialsearch.

We did not use methodological filters in database searches to avoid possible omission of relevant studies, according to the recommendations of Leeflang et al. [13]. The search terms included and captured the concepts of ‘endometrial,’ “cancer,” “carcinoma,” ‘transvaginal ultrasound,’ ‘sonography,’ ‘myometrial,’ and ‘magnetic resonance imaging.’ Language restriction in the search was set to English.

One author (JLA) screened the titles and abstracts identified by the searches to exclude obviously irrelevant article, i.e., those not strictly related to the topic under review. Full-text articles were obtained to identify potentially eligible studies, and 3 reviewers (JLA, RS, BN) applied independently the following inclusion criteria:

To avoid inclusion of duplicate cohorts in the meta-analysis in the case of 2 studies from the same authors, the study period of each study was examined; if dates overlapped, we chose the latest study according to the publication date, considering that patients from the first study were also included in the latest one. We used “snowball” strategy to identify potential interesting papers by reading reference list of those papers selected for full text reading. No attempts were made to contact the authors.

The Patients, Intervention, Comparator, Outcomes, Study design (PICOS) criteria were used for describing the studies included.

Diagnostic accuracy results and additional useful information on patients and procedures were retrieved from selected primary studies independently by 3 of the authors (JLA, RS, BN). Disagreements arising during the process of study selection and data collection were resolved by consensus among 2 of the authors (JLA, BN).

Quality assessment was conducted, adapting to this particular review the tool provided by the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) [14]. The QUADAS-2 format includes 4 domains: 1) patient selection, 2) index test, 3) reference standard, and 4) flow and timing. For each domain, the risk of bias and concerns about applicability (the latter not applying to the domain of flow and timing) were analyzed and rated as low, high or unclear risk. The results of quality assessment were used for descriptive purposes to provide an evaluation of the overall quality of the included studies and to investigate potential sources of heterogeneity. Two authors (JLA, BN) evaluated independently the methodological quality, using a standard form with quality assessment criteria and a flow diagram; they resolved disagreements by discussion between 2 of the authors.

The methodology of quality criteria was based on description of inclusion and exclusion criteria for patient selection domain, description about how the of the index test (TVS/MRI) was performed and interpreted for index test domain, description of reference standard used and whether pathologists were blinded or not to index test for reference standard domain and description of the time elapsed from index test assessment to reference standard result.

We extracted or derived information on diagnostic performance of TVS and MRI. A random-effects model was used to determine overall pooled sensitivity, specificity, positive likelihood ratio (LR+) and negative likelihood ratio (LR−). Positive and negative likelihood ratios (LRs) were used to characterize the clinical utility of a test and to estimate the post-test probability of disease. A LR of 0.2–5.0 provides weak evidence for either ruling out or confirming the disease. A LR of 5.0–10.0 and 0.1–0.2 provides moderate evidence to either confirm or rule out the disease. A LR >10 or <0.1 provides strong evidence to either confirm or rule out the disease [15].

Using the mean prevalence of deep (>50%) MI (pretest probability) in each subset, depending upon the technique assessed and LRs, post-test probabilities were calculated and plotted on Fagan nomograms.

We assessed the presence of heterogeneity for sensitivity and specificity using Cochran's Q statistic and the I² index [16]. A p-value<0.1 indicates heterogeneity. The I² index describes the percentage of total variation across studies that is due to heterogeneity rather than chance. According to Higgins et al. [16], I² values of 25%, 50%, and 75% would be considered to indicate low, moderate and high heterogeneity, respectively. Forest plots of sensitivity and specificity of all studies were plotted.

Summary receiver-operating characteristics (sROC) curves were plotted to illustrate the relationship between sensitivity and specificity. Comparison of diagnostic performance between TVS and MRI for detecting deep MI was done using the bivariate method [15]. Meta-regression was used if heterogeneity existed to assess covariates that could explain this heterogeneity. The covariates analyzed were sample size, prevalence, mean patient age and number of observers (single/multiple). Publication bias was assessed by a regression of diagnostic log odds ratio against 1/√(effective sample size), weighted by effective sample size, with p<0.10 for the slope coefficient indicating significant asymmetry [17].

All analyses were performed using Meta-analytical Integration of Diagnostic Accuracy Studies (MIDAS) and METANDI commands in STATA version 12.0 for Windows (Stata Corporation, College Station, TX, USA). A p-value<0.05 was considered as statistically significant.

The electronic search provided a total of 747 citations. We did find 19 citations in websites for ongoing trials, but none of them were studies specifically related to the topic of the present meta-analysis. After removal of 371 duplicate records and 33 papers published in language other than English, 343 citations remained. Of these, 325 were excluded because it was clear from the title or abstract that they were not relevant to the review (papers assessing TVS but not MRI [n=54], papers assessing MRI but not TVS [n=124], papers not assessing diagnostic performance or not related to the topic [n=92], reviews [n=54] or letters to the editor [n=1]). We examined the full text of the remaining 18 articles. Finally, 10 studies were discarded because they did not meet inclusion criteria (studies using 3D ultrasound [n=5], studies using transabdominal ultrasound [n=2], studies in which data as table 2×2 was not possible to obtain [n=2], and one study that analyzed TVS and MRI but in different sets of patients). Thus, the remaining 8 studies were included in the review and meta-analysis [1819202122232425]. No additional relevant studies were found from references cited in the papers included in the review. A flowchart summarizing literature identification and selection is given in Fig. 1.

A total of 8 studies [1819202122232425] published between January 1992 and February 2013 reporting on 560 patients were included in the final analyses. Among these 560 women, 192 had deep MI. Mean prevalence of deep MI was 33.7%, ranging from 7.1% to 52.4% [1819202122232425]. All studies reported the clinical characteristics of the cohort to some extent. Mean patients' age was reported in 6 out of 8 studies and ranged from 55 to 65 years. Table 1 shows PICOS features of the studies included.

Study design was clearly stated as prospective in 6 studies [202122232425]. In 2 studies, design was not clear [1819]. A graphical display of the evaluation of the risk of bias and concerns regarding applicability of the selected studies is shown in Fig. 2.

Regarding risk of bias and the domain patient selection, 1 study was not clear regarding patient inclusion criteria [18] and 3 were considered as high risk for patient selection since only patients with “conclusive” or unequivocal” results for TVS and/or MRI were included [232425].

Concerning the domain index test, with regard to TVS 5 studies adequately described the method of index text as well as how it was performed and interpreted, 2 studies were unclear [1822] and one was considered as high risk since MI was estimated “subjectively” by the examiner [25]. With regard MRI, 5 studies adequately described the method of index text as well as how it was performed and interpreted, 3 studies were unclear [182225].

Concerning the domain flow and timing, the time elapsed between the index test and reference standard was unclear in 2 studies [2223].

For the domain reference standard, all studies were likely to correctly classify the target condition by the reference standard. However, in 3 studies it was not clearly specified if the results of the reference standard were interpreted using gross evaluation of the uterus or permanent frozen section [212325]. Only 2 studies [1819] reported specifically that pathologists were blinded to imaging results, in the rest of the studies this was unclear.

Regarding applicability, for the domain patient selection, all studies were deemed to include patients that matched the review question. For the domain index test, most studies were considered as having low concerns for applicability as the index tests were described well enough for study replication, as was the reference standard domain.

Overall, pooled sensitivity, specificity, LR+, and LR− of TVS for detecting deep MI were 75% (95% confidence interval [CI]=67%–82%), 86% (95% CI=75%–93%), 5.6 (95% CI=3.0–10.2), and 0.28 (95% CI=0.22–0.37), respectively. Low heterogeneity was found for sensitivity (I²=24.2%; Cochran Q=9.24; p=0.240) but significant heterogeneity was found for specificity (I²=80.6%; Cochran Q=36.14; p<0.001). On the other hand, pooled sensitivity, specificity, LR+, and LR− of MRI for detecting deep MI were 83% (95% CI=76%–89%), 82% (95% CI=72%–89%), 4.7 (95% CI=3.0–7.2), and 0.20 (95% CI=0.14–0.29), respectively. Low heterogeneity was found for sensitivity (I²=5.7%; Cochran Q=7.42; p=0.390) but significant heterogeneity was found for specificity (I²=83.4%; Cochran Q=42.21; p<0.001). No statistical differences were found when comparing both methods (p=0.314).

Fig. 3 shows forest plots for both methods. It can be observed in these graphics that most studies had acceptable CIs for both sensitivity and specificity for MRI and TVS. sROC curves are shown in Fig. 4. It can be observed that both techniques had similar areas under the curve for sROC curves, but the 95% prediction contour is narrower for MRI as compared with TVS. Fagan nomograms show that a positive test for TVS and MRI increases significantly the pretest probability deep MI, from 34% to 74% in case of TVS and from 34% to 71% in case of MRI, while a negative test significantly decreases the pretest probability, from 34% to 13% in case of TVS and from 34% to 9% in case of MRI (Fig. 5). Meta-regression showed that sample size, prevalence, mean patient age, and number of observers (single/multiple) did not explain heterogeneity observed for specificity. No publication bias was found, neither for TVS (p=0.650) nor for MRI (p=0.090).