Contrast-enhanced ultrasound for the characterization of portal vein thrombosis vs tumor-in-vein in HCC patients: a systematic review and meta-analysis

This study was performed following the guideline of the Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy. We searched four databases, i.e., PubMed, Embase, Cochrane Library, and Web of Science, for articles published up to the 5th of May 2019. The search strategies were the combination of MeSH terms, entry terms (synonyms), study keywords, and search filters for diagnostic tests [22]. The search terms included (“portal vein thromb*” OR “portal vein embolus” OR “PVT” OR (“tumor in vein” AND “portal vein”)) AND (“contrast enhanced ultraso*” OR “CEUS” OR “contrast enhanced sonography” OR “contrast-enhanced ultraso*” OR “contrast-enhanced doppler ultrasonography”).

Two reviewers (Chen and Zhu) independently screened titles and abstracts, and discrepancies were eliminated via discussion. The articles were carefully reviewed with the following eligibility criteria:

Two reviewers (Chen and Zhu) independently extracted the data and discrepancy was eliminated by discussion. Study characteristics, patients’ features, and data of diagnostic results (e.g., TP, TN, FP, FN, sensitivity, and specificity) of each included study were extracted.

Two other reviewers (Zhang and Song) independently assessed the quality of each study using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool and disagreement was resolved by discussion.

Data was analyzed using the Stata 15.0 software (Midas commands) (StataCorp LP, College Centre) and Review Manager (RevMan) 5.3 (The Nordic Cochrane Centre, The Cochrane Collaboration).

For each study, extracted information was used to construct two × two contingency tables. The bivariate random effect model based on the binomial distribution for sensitivity and specificity was performed. To determine whether a threshold effect is present, we applied the spearman correlation analysis with p < 0.05 representing a threshold effect. The pooled sensitivity and specificity were displayed with pooled point estimate and 95% confidence intervals (CIs). The hierarchical summary receiver-operating characteristic (ROC) curve was plotted and the area under the curve (AUC) was calculated. The Fagan plot was performed to estimate how much probability that a patient suffers from TIV would change by CEUS [23].

The I² index was calculated to estimate the heterogeneity, which shows the total variation (in percentage) across studies due to heterogeneity rather than chance. The overall studies demonstrated substantial heterogeneity if the I² index value is greater than 50% and vice versa. Covariates that may contribute to heterogeneity were evaluated by subgroup analysis to figure out the origin of heterogeneity.

Subgroup analysis was conducted in consideration of study scale in regard to enrolled subject numbers, country, study design (prospective or retrospective), blind method, specific CEUS contrast agents, and diagnostic criteria if based on pathology. Sensitivity analysis was also performed to address quality differences and validate the robustness of our result.

To investigate the publication bias, Deeks’ funnel plot with a regression line was plotted [24]. The slope coefficient is suggestive of an asymmetric funnel plot if p < 0.10 [25].

The initial search from four databases identified 986 articles, of which 312 were removed because of duplication and 560 were removed after reading the title and abstract carefully. Besides, 80 articles were excluded due to article type, of which eight were reviews and 72 were conference abstracts. Twenty non-English language articles were removed from the screen list of full-text assessment.

After reading fourteen articles in the full-text assessment list, two of them failed to provide sufficient data, three of them enrolled unsuitable patients, who might cause bias if enrolled, and one had no reference standard. One was removed due to duplicate patients. Finally, seven articles were included in this systematic review and meta-analysis (Fig. 1, Table 1).

QUADAS-2 was selected as the tool for evaluating the included studies. This consists of two parts to assess the study quality: the risk of bias part contains four categories (patient selection, index test, reference standard, and flow and timing) and the applicability concerns part contains three categories (patient selection, index test, and reference standard). The risk grades were assigned as high, unclear, and low by two independent reviewers (Zhu and Song) and the discrepancies were settled via discussion. All seven included studies demonstrated a low risk of bias and no applicability concerns in patient selection (Fig. 2). One study [26] has an unclear risk of bias and high applicability concern in index test, because contrast-enhanced color Doppler imaging was adopted rather than contrast-enhanced ultrasound imaging to characterize TIV from PVT, which might cause bias and limit the applicability. About 42.8% of the studies [26‐28] were graded as unclear risk of bias in reference standard because these studies did not report whether the application of reference standard was in a blind feature. Some studies applied pathological results as the only reference standard, while some applied a mixture of pathology and imaging follow-up; therefore, 58.2% (4/7) of the studies [26‐29] were graded as high risk of bias in the category of flow and timing. In summary, using the QUADAS-2 tool, the included studies have a low concern regarding applicability in all the three categories and low risk of bias in patient selection and index test categories, but might be at risk in the categories of reference standard and flow and timing.

Among the seven included studies, no threshold effect was detected (p = 0.590). The summary sensitivity and specificity of CEUS in diagnosing TIV were 0.94 (95%CI, 0.89–0.97) and 0.99 (95%CI, 0.80–1.00) respectively. After plotting the summary ROC (SROC) curve, the calculated value of AUC of SROC was 0.97 (95%CI, 0.95–0.98) (Fig. 3, supplementary Table 1).

According to the forest plot (Fig. 4), the I² of summary sensitivity and specificity is 27.2% (p = 0.22) and 77.08% (p < 0.01). One study [28] showed obvious heterogeneity from the others in specificity, because only 3 patients without TIV were enrolled in this study.

Due to the limited number of false-positive subjects (three in total) in our review, some subgroup analysis and sensitivity analysis results were difficult to obtain due to the convergence problem. The sensitivity and AUC were consistent across all the subgroups, whereas the 95%CI of the specificity within subgroups of study scale, country, and diagnostic criteria varied a lot when compared with the overall result (Table 2). Except for Song’s study [28] with a lower 95%CI of specificity of 0.51, the sensitivity analysis indicated little quality difference among the included studies (supplementary Table 2).

Publication bias was not detected by performing Deeks’ funnel plot in this study. Besides, the slope coefficient indicated no significant small sample size bias (Fig. 5, p = 0.54).

The Fagan plot (supplementary Fig. 1) demonstrated that CEUS is very informative in raising the probability of diagnosing TIV from 50 to 99% when positive and lowering the probability of malignancy to as low as 5% when negative.