Diagnostic accuracy of EUS and CT of vascular invasion in pancreatic cancer: a systematic review

Abstract

Objective

The objective of this study was to summarize the accuracy of preoperative vascular invasion with endoscopic ultrasound (EUS) and computed tomography (CT) test performance in pancreatic cancer with meta-analysis

Method

Two reviewers searched MEDLINE database to identify relevant studies. The reference lists of the trials were manually searched. Included studies used surgical and/or histological findings as the “gold standard,” and provided sufficient data to construct a diagnostic 2 × 2 table. A statistical program of Meta-Disc was used to calculate the pooled sensitivity, specificity, positive LR, negative LR, DOR, and the SROC curve. Publication bias was assessed by Deeks’ asymmetry test. Sensitivity analysis and subgroup analysis were calculated to down the heterogeneity. Meta-regression was calculated to evaluate potential sources of heterogeneity

Result

A total of 30 studies with 1,554 patients were included for the analysis, nine of these studies compared EUS with CT to assess the diagnostic efficiency The pooled sensitivity of EUS and CT was 72 % (95 % CI 67–77 %) and 63 % (95 % CI 58–67 %), and the pooled specificity of EUS and CT was 89 % (95 % CI 86–92 %) and 92 % (95 % CI 90–94 %), respectively. The positive LR of EUS and CT was 5.14 (95 % CI 3.14–8.40) and 6.21 (95 % CI 3.96–9.71), and the negative LR was 0.36 (95 % CI 0.25–0.52) and 0.41 (95 % CI 0.31–0.55), respectively. The AUCs of EUS and CT were 0.9037 and 0.8948. The subgroup analysis of nine studies performed both EUS and CT showed CT scan with a lower sensitivity of 48 % (95 % CI 0.40–0.56), when compared to EUS of 69 % (95 % CI 0.61–0.77). The overall AUCs of CT scan appear to be lower (AUCs = 0.8589), compared with EUS (AUCs = 0.9379)

Conclusion

EUS performed better than CT in differentiating vascular invasion preoperative on pancreatic cancer. EUS could provide other additional information when compared with CT.

Introduction

Pancreatic cancer was a virulent disease with a poor prognosis and a leading cause of cancer death worldwide (James and Gibbs 2005; Matsuda et al. 2008). Although recent advancement in surgical management and adjuvant treatment remained the potentially curative therapy, only about 10 % of pancreatic cancer could undergo curative surgery at the time of diagnosis, and overall, 1-year survival and 5-year survival were <1 and 5 %, respectively(Ghaneh et al. 2007; Jemal et al. 2009). Therefore, it was important for us to identify patients who were likely to benefit from surgical resection. Besides metastatic disease precluding curative therapy, assessment of vascular invasion was considered to preclude surgical resection (Hackert and Büchler 2013; Siriwardana and Siriwardena 2006) and was also an important predictor for poor prognosis (Ravikumar et al. 2014; Foroughi et al. 2012). As a result, an accurate preoperative evaluation of vascular invasion was crucial in determining treatment modalities.

Our previous study has found that EUS-FNA performed good in diagnosing pancreatic cancer (Chen et al. 2012). Besides FNA for tissue diagnosis, EUS could offer high-resolution, local imaging of the pancreas and surrounding structures. During the last 20 years, CT has been used for vascular invasion because of its ability to noninvasively perform. So far, there remains controversial on which was the most optimal imaging tool to decide vascular invasion in pancreatic cancer. The aim of this meta-analysis is to compare EUS with CT in preoperative evaluation of vascular invasion in patients with pancreatic cancer.

Materials and methods

Search strategy

We identified all studies evaluating the use of computed tomography and/or EUS on diagnostic performance of vascular invasion in pancreatic cancers, and we used the guidelines to conduct the systematic review (Leeflang et al. 2008; Devillé et al. 2002). A literature search was performed using MEDLINE database. The search strategy was based on combinations of the following terms: vascular invasion, vessel involvement, vessel, pancreatic cancer, pancreatic carcinoma, pancreatic neoplasm, endoscopic ultrasound, EUS, and computed tomography, CT. The restrictions included English language publications and human subject. Other database such as ELSEVIER Science Direct (SDOS), EBSCO, and Springer Link was also searched for relevant articles. The reference lists of the trials were also manually searched.

Study selection criteria

Articles in which vascular invasion were diagnosed by EUS and/or CT were selected. Inclusion criteria included in the analysis were as follows:(1) data from the selected studies were provided accurately and confirmed by surgical and/or histological findings; (2) we collect retrospective or prospective clinical trial; and (3) only studies from which a 2 × 2 table could be constructed (actual counts for the number of false positive (FP), false negative (FN), true positive (TP), and true negative (TN) were available). Studies were excluded if anyone of the inclusion criteria was not met. Case reports, letters, descriptive reviews, and articles published in languages other than English were excluded.

Data extraction

Two independent investigators checked the titles and abstracts of the potential studies. From the eligible studies, we extracted relevant data from the final full text. To perform accuracy analyses, we extracted the following items: study characteristics, including author, year of publication, country, sample size, retrospective or prospective design, type of EUS and CT, and confirmatory procedure; the number of true positive, true negative, false negative, and false negative. With extracted data, 2 × 2 tables were constructed.

Quality of studies

Methodological quality was assessed using the following criteria based on the STARD statement (Bossuyt et al. 2003). The same two investigators independently utilized components of this statement to judge the quality of literatures including: a consecutive series of participants; prospective or retrospective study; sufficient description or standardization of technique to enable replication; independent comparison with reference standard; blind interpretation; reporting how indeterminate results; and missing responses of the index tests were handled. This was a 6-point quality scale, and the statement defined low-quality studies having a score of <3 and high-quality studies having a score ≥3.

Statistical analysis

Based on the 2 × 2 table for each study, we extracted TP, FP, TN, and FN values and entered these into the statistical program Meta-Disc version 1.4 (Zamora et al. 2006) (Meta-Disc, unit of clinical biostatistics team of the Romanycajal hospital, Madrid, Spain). The following measure of test accuracy of EUS and/or CT was done by calculating pooled estimates of sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR),and diagnostic odds ratio(DOR) with corresponding 95 % CIs. Mantel–Haenszel method (fixed effects model) or the DerSimonian–Laird method (random effects model) was applied to pool the data of individual studies. And we also performed sensitivity analysis and calculated sensitivity, specificity, LR+, and LR− for subgroup of high-quality studies.

We used the chi-square (χ ²) and inconsistency index (I ²) to assess heterogeneity across the studies (Higgins et al. 2003). In the chi-square test, the p < 0.05 was considered having significant heterogeneity among the studies. An inconsistency index of 0–20 % indicates homogeneity, and values >50 % indicate high heterogeneity. We constructed a summary receiver operating characteristic (SROC) curve for the diagnostic performance of the sensitivity and specificity. The area under the curve (AUC) was used as a summary for the diagnostic accuracy. High test power had an AUC close to 1, and a poor power had an AUC close to 0.5. Threshold analysis was performed using the Spearman coefficient.

If heterogeneity existed, a random effect model was applied to obtain a summary estimate for sensitivity and specificity. We performed subgroup analysis to assess potential sources of heterogeneity. Studies were allocated to prespecified subgroups according to STARD score, year of publication, and head-to-head comparison. We also performed meta-regression to evaluate potential heterogeneity. The following potential sources were included: the study design year of publication, sample size, blind interpretation, and linear EUS or not. The results of the meta-regression were presented as relative diagnostic odds ratio (RDOR) corresponding covariate. Publication bias was assessed by Deeks’ asymmetry test and a funnel plots by using diagnostic log odds ratio versus 1/sqrt (effective sample size), with p < 0.05 indicating significant asymmetry.

Result

Through the above-mentioned search strategy, a total of 933 publications were identified. Of these, 71 relevant articles were extracted for a full-text reading and critical assessment. Twenty-seven studies that met the inclusion criteria were finally included in this meta-analysis. Addition, three other studies were included in this review after the reference lists of the trials were manually searched.

Among the included studies, nine studies performed EUS in diagnosing vascular invasion, and 12 studies performed CT in diagnostic test; nine studies compared EUS with CT to assess the diagnostic efficiency. Twelve studies reported consecutive series of patients; 18 of the 30 studies had a prospective study design. The methodological quality was not good in all included studies, and the quality score was showed in Table 1. Six studies included data on ampullary tumors and other subtypes, which might lead to heterogeneous. Altogether, 30 studies reported 1554 patients were performed EUS and/or CT diagnosis of vascular invasion with pancreatic cancer.

Table 1.

Characteristics of the included articles

Study (year)	Country	Method	Sample (EUS/CT)	Confirmatory procedure	EUS/CT technique
Seicean et al. (2008)	Romania	EUS	30	Surgery	GF-UM160
Aslanian et al. (2005)	USA	EUS	30	Surgery	GF-UM20, GF-UM130, GF-UC30P
Fritscher-Ravens et al. (2005)	UK	EUS	22	Surgery	FG-32UA, FG-34UX, EUB 8000, EUB 6000
Rosch et al. (2000)	Germany	EUS	75	Surgery, angiography, and pathology	GF-UM3, UM-20
Yusoff et al. (2003)	Australia	EUS	32	Surgery and histology	GF-UM3, GF-UM20, or GF-UM 130
Yasuda et al. (1993)	Japan	EUS	29	Surgery	GF-UM2, GF-UM3
Snady et al. (1994)	USA	EUS	38	Surgery	GF-UM2 7.5-MHz
Buscail et al. (1999)	France	EUS	32	Surgery	EU-M3, EU-M20
Bao et al. (2008)	USA	EUS	27	Surgery	ND
Buchs et al. (2007)	Switzerland	EUS/CT	90/153	Surgery and pathology	GIF-EUM-20, GIFEUM 30 and 160/ GE high-speed CT, MX-8000
Mertz et al. (2000)	USA	EUS/CT	16/16	Surgery	FG 32UA linear array/ helical CT¤
Ramsay et al. (2004)	Australia	EUS/CT	19/19	Surgery	GF-UM3, GF-UM20 or GF-UM 130/ a single-array spiral CT¤
Soriano et al. (2004)	Spain	EUS/CT	62/60	Surgery	GF-UM20/ helical CT¤
Gress et al. (1999)	USA	EUS/CT	75/58	Surgery	EUM-20FG 32UA/ helical CT¤
Tellez-Avila et al. (2012)	France	EUS/CT	40/50	Surgery and pathology	GF-UCT140/ 16- or 64-slice multi-detector CT
Midwinter et al. (1999)	UK	EUS/CT	31/30	Surgery	GF-UM 20 radial scanning/ helical CT¤
Rivadeneira et al. (2003)	USA	EUS/CT	44/44	Surgery and pathology	FG32UA/ helical CT¤
Shoup et al. (2000)	USA	EUS/CT	37/37	Surgery	UM-30/ axial or helical CT¤
Klauss et al. (2008)	Germany	CT	28	Surgery and pathology	Siemens Somatom Sensation 16¤
Vellet et al. (1992)	Canada	CT	39	Surgery and pathology	9800 Quick (GE)
Brügel et al. (2004)	Germany	CT	33	Surgery and pathology	ND
Furukawa et al. (1998)	Japan	CT	27	Surgery	a 9,500 or X-vigor scanner
Lepanto et al. (2002)	Canada	CT	36	Surgery	ND
Megibow et al. (1995)	USA	CT	118	Surgery and pathology	9,800 HiLight or HiLight advantage
Diehl et al. (1998)	Germany	CT	76	Surgery	Helical CT¤
Arslan et al. (2001)	Norway	CT	48	Surgery	Helical CT¤
Ichikawa et al. (1997)	Japan	CT	21	Surgery	X-Force SH and TCT900S
Karmazanovsky et al. (2005)	Russia	CT	65	Surgery and pathology	spiral CT¤
Li et al. (2012)	China	CT	18	Surgery	multi-detector row CT scanner¤
Lee et al. (2010)	Korea	CT	222	Surgery and pathology	1.5 T, Siemens, 3-mm section, dual phase

ND not defined, ¤ represent no specific details

Diagnostic accuracy

Figure 1 showed the forest plots of pooled sensitivity of EUS and CT were 72 % (95 % CI 67–77 %) and 63 % (95 % CI 58–67 %), respectively. From the forest, the sensitivity of EUS from various studies ranged from 20 to 100 %, and CT ranged from 13 to 93 %. Figure 2 showed a pooled specificity of EUS and CT was 89 % (95 % CI 86–92 %) and 92 % (95 % CI 90–94 %), respectively. The tests of heterogeneity were all highly significant (χ ² p < 0.001, I ² > 50 %). Therefore, we used the random effects model for pooled estimate. The positive LR of vascular invasion using EUS and CT for preoperative staging was 5.14 (95 % CI 3.14–8.40) and 6.21 (95 % CI 3.96–9.71) (Fig. 3), and the negative LR was 0.36 (95 % CI 0.25–0.52) and 0.41 (95 % CI 0.31–0.55) (Fig. 4), respectively.

Fig. 1.

a Forest plot of pooled sensitivity of EUS in detecting vascular invasion; b forest plot of pooled sensitivity of CT in detecting vascular invasion. The size of each point is proportional to the sample size for each study, and the horizontal lines through the points indicate a graphical representation of the 95 % CI of that study

Fig. 2.

a Forest plot of pooled specificity of EUS in detecting vascular invasion; b forest plot of pooled specificity of CT in detecting vascular invasion. The size of each point is proportional to the sample size for each study, and the horizontal lines through the points indicate a graphical representation of the 95 % CI of that study

Fig. 3.

a Forest plot of pooled positivity LR of EUS in detecting vascular invasion; b forest plot of pooled positivity LR of CT in detecting vascular invasion. The size of each point is proportional to the sample size for each study, and the horizontal lines through the points indicate a graphical representation of the 95 % CI of that study

Fig. 4.

a Forest plot of pooled negativity LR of EUS in detecting vascular invasion; b forest plot of pooled negativity LR of CT in detecting vascular invasion. The size of each point is proportional to the sample size for each study, and the horizontal lines through the points indicate a graphical representation of the 95 % CI of that study

The pooled diagnostic odds ratio (DOR), the odds of having vascular invasion in positive as compared to negative EUS studies, was 23.44 (95 % CI 11.54 to 47.57), with a high statistical heterogeneity (χ ² = 29.75, p = 0.03, I ² = 42.9 %). The pooled diagnostic odds ratio of CT was 20.07 (95 % CI 10.32–39.05), with p < 0.001, and I ² index was 60.8 %. For EUS diagnosis, the area under the SROC curves (AUCs) and Q* index were 0.9037 and 0.8352, respectively, indicating high test accuracy (Fig. 5). For CT diagnosis, the area under the SROC curves (AUCs) and Q* index were 0.8948 and 0.8257 (Fig. 5), respectively. Both modalities offered a similarly good diagnostic accuracy for vascular invasion. The SROC curves did not show “Shoulder arm” like distribution, with Spearman correlation coefficient of EUS was 0.149 (p = 0.556), and CT was 0.410 (p = 0.065).

Fig. 5.

a Summary receiver operating characteristic (SROC) curve of all 18 studies with EUS test. b Summary receiver operating characteristic (SROC) curve of all 21 studies with CT test

Subgroup analysis and sensitivity analysis

In sensitivity analysis for EUS test, we included nine high-quality studies (Table 2). The pooled sensitivity was 70 %, and the pooled specificity was 92 %. The positive LR was 6.24, and the negative LR was 0.37. The DOR was 24.04, and AUCs was 0.9128. The inconsistency index for heterogeneity decreased substantially, from 42.09 to 19.8 %, and the p value for heterogeneity test was more than 0.05. In sensitivity analysis for CT test, we included 14 high-quality studies and excluded Megibow et al. study. Megibow et al. study could be seen as a relative outlier, in which the result showed lower sensitivity and specificity. The pooled sensitivity, specificity, positive LR, negative LR, DOR, and AUCs were 63 %, 93 %, 6.26, 0.44, 20.33, and 0.9263, respectively. The inconsistency index for heterogeneity also decreased substantially.

Table 2.

Overall, sensitivity analysis results and a head-to-head analysis

	No. of studies	SEN	SPE	PLR	NLR	DOR	AUC
Sensitivity analysis (with a score above 3)
EUS	9	70 % (0.61–0.77)	92 % (0.87–0.95)	6.24 (3.81–10.21)	0.37 (0.25–0.55)	24.04 (11.31–51.10)	0.9128
P value and I 2		0 and 68.7 %	0.074 and 44.2 %	0.283 and 17.9 %	0.01 and 59.7 %	0.267 and 19.8 %
CT (Megibow et al. study exclude)	14	60 % (0.53–0.66)	93 % (0.91–0.96)	6.26 (4.12–9.52)	0.44 (0.31–0.63)	20.33 (10.24–1.27)	0.9263
P value and I 2		0 and 81.3 %	0.085 and 36.3 %	0.357 and 8.8 %	0 and 86.0 %	0.095 and 35.1 %
“Head-to-head” analysis
EUS	9	69 % (0.61–0.77)	94 % (0.90–0.97)	7.62 (4.58–12.68)	0.38 (0.24–0.63)	29.37(11.56–74.62)	0.9379
P value and I 2		0 and 77.9 %	0.043 and 49.9 %	0.377 and 7.0 %	0 and 74.9 %	0.101 and 40 %
CT	9	48 % (0.40–0.56)	93 % (0.89–0.95)	5.59 (3.65–8.57)	0.53 (0.39–0.73)	12.57 (6.85–23.07)	0.8589
P value and I 2		0.003 and 65.7 %	0.012 and 59.3 %	0.612 and 0	0 and 71.6 %	0.670 and 0

A subgroup analysis of nine studies compared EUS with CT to assess the diagnostic efficiency. As shown in Table 2, CT scan showed a lower sensitivity of 48 % (95 % CI 0.40–0.56), when compared to EUS of 69 % (95 % CI 0.61–0.77). The specificity of CT scan and EUS was 93 % (95 % CI 0.89–0.95) and 93 % (95 % CI 0.89–0.95), respectively. The overall AUCs of CT scan appear to be lower (AUCs = 0.8589), compared with EUS (AUCs = 0.9379). However, when we estimated the specificity, positive LR, DOR, and SROCs, the CT scan appeared to be outstanding in homogeneity (p > 0.05, I ² < 50 %).

To standardize the change technology and the change criteria of EUS and/or CT for vascular invasion, the studies were group into three time periods. They were 1993–2009, 2000–2005, and 2006–2013. All pooled estimates during the three time periods were given in Table 3. For the more recent time periods, EUS had a lower pooled sensitivity of 73 % (95 % CI 61–83 %) and specificity of 84 % (95 % CI 76–90 %), respectively, compared with the earlier time periods. CT had a better accurate for vascular invasion, with the AUC 0.9769.

Table 3.

Pooled diagnostic accuracy of EUS and CT for different periods (95 % CIs)

	NO. of studies	Sen (%)	Spe (%)	PLR	PLR	DOR	AUC
EUS
1993–1999	6	87 % (0.80–0.93)	95 % (0.89–0.98)	8.39 (2.88–24.41)	0.18 (0.06–0.60)	71.38 (21.45–237.57)	0.9563
2000–2005	8	58 % (0.48–0.67)	89 % (0.83–0.93)	5.9 (2.37–14.67)	0.49 (0.33–0.74)	17.75 (5.19–60.70)	0.8687
2006–2013	4	73 % (0.61–0.83)	84 % (0.76–0.90)	3.25 (1.84–.75)	0.42 (0.29–0.63)	12.23 (5.40–27.72)	0.8539
CT
1993–1999	7	56 % (0.59–0.63)	90 % (0.84–0.94)	5.73 (1.88–17.47)	0.47 (0.28–0.78)	17.78 (3.27–96.79)	0.8974
2000–2005	9	74 % (0.65–0.81)	90 % (0.86–0.94)	5.14 (3.17–8.33)	0.37 (0.24–0.56)	20.70 (10.36–4.59)	0.8911
2006–2013	5	60 % (0.51–0.69)	94 % (0.91–0.96)	9.02 (4.14–19.65)	0.43 (0.29–0.65)	22.21 (7.07–69.77)	0.9769

Meta-regression and publication bias

We performed a meta-regression by identified six potential sources: (1)consecutive; (2)study design (prospective versus retrospective); (3) sample size (<40 vs ≥40); (4) year of publication (<2002 vs ≥2002); (5) blind or not; and (6) linear EUS or not. The outcomes of the regression were displayed in Table 4. The meta-regression did not show any statistical significance.

Table 4.

Meta-regression analysis to determine sources of heterogeneity for all 30 studies

Var	Coefficient	p value	RDOR	(95 % CI)
EUS
Blind	−0.576	0.5828	0.56	(0.06; 5.27)
Sample	0.523	0.5944	1.69	(0.21; 13.77)
Design	0.397	0.7795	1.49	(0.07; 31.30)
Consecutive	−0.418	0.7136	0.66	(0.06; 7.57)
Year	−1.397	0.1540	0.25	(0.03; 1.84)
Linear	0.854	0.5033	2.35	(0.15; 36.43)
CT
Design	−0.648	0.3834	0.52	(0.11; 2.40)
Year	0.036	0.9638	1.04	(0.20; 5.42)
Sample	0.550	0.4692	1.73	(0.36; 8.40)
Consecutive	1.493	0.1105	4.45	(0.68; 29.08)

There was a method used for detecting small study effects based on a linear regression of log odds ratios on effective sample sizes. The Deek’s funnel plot asymmetry test showed slope was not significant, suggesting no major publication bias (Fig. 6).

Fig. 6.

Results of Deeks’ funnel plot asymmetry test of EUS (a) and CT (b) for publication bias. The nonsignificant slope indicates that no significant bias was found. ESS effective sample size

Discussion

A variety of studies have evaluated the diagnostic efficiency of various imaging techniques on vascular invasion in pancreatic cancer. CT was widely used on preoperative evaluation of vascular invasion for its ability to noninvasively perform. However, several studies demonstrated a controversial accuracy that ranged from 55 to 91 % (Megibow et al. 1995; Li et al. 2012). Similarly, endoscopic ultrasound has been recognized as a useful and accurate modality on preoperative evaluation for the past 10 years and has been used to determine the resectability of pancreatic cancer. Several studies have suggested that EUS was superior to CT in the detection vascular invasion (Buchs et al. 2007; Gress et al. 1999). EUS was a dynamic study much more dependent than CT on operator experience as well as patient factors that may make a study difficult. Recent comparative studies presented more varied and conflicting data when determining vascular invasion (Yusoff et al. 2003; Soriano et al. 2004).

Our meta-analysis systematically summarized the ability of EUS on preoperative evaluation and reported herein showed that EUS has an excellent accuracy on predicting vascular invasion, with an area under the SROCs curve of 0.9037. The meta-analysis showed a pooled sensitivity of 72 % and a pooled specificity of 89 %, respectively. However, a weakness of this meta-analysis was that there was high heterogeneity (p < 0.05, I ² > 50 %). To correct for potential sources of heterogeneity, we perform sensitivity analysis and meta-regression. Although the outcomes of the meta-regression did not show any statistical significance, the sensitivity analysis showed the inconsistency index for heterogeneity decreased substantially. The diagnosis odds ratio (DOR) was 24.04, which means that, if EUS indicated vascular invasion, the patient had odds of 24.04 times to have true anatomic vascular invasion. In addition, this meta-analysis showed CT had a pooled sensitivity of 62 % and a pooled specificity of 92 %. When we excluded Megibow et al. (1995) study from the sensitivity analysis, the result showed a sensitivity of 60 % and a specificity of 93 %, with the heterogeneity decreased significantly. And the diagnosis odd was 20.33.

However, differences in study designs between the EUS and CT studies could have influenced the pooled estimates. Based on our head-to-head comparison, we perform a subgroup analysis of nine studies with both CT and EUS in the same study as a preoperative diagnostic tool to minimize the population bias. Our results showed that EUS appeared to be more sensitive in the diagnosis of vascular invasion (SEN 69 % vs. SEN 48 %). Both EUS and CT revealed similar specificities, and we could confirm the higher performance of EUS over CT on a patient bias (AUC 0.9379 for EUS vs. AUC 0.8589 for CT).

The likelihood ratio (LR) was a measure of how good a diagnostic test was and used to help in selecting an appropriate diagnostic test. They not only could reflect the sensitivity and specificity, but also were more stable than sensitivity and specificity. In the present study, the positive LR of vascular invasion using EUS and CT was 5.14 (95 % CI 3.14–8.40) and 6.21 (95 % CI 3.96–9.70), and the negative LR was 0.36 (95 % CI 0.25–0.52) and 0.41 (95 % CI 0.31–0.55), respectively, showing that both EUS and CT had similar accurate in predicting preoperative evaluation of vascular invasion.

Five of the 18 studies reported linear-array EUS was utilized in isolation, either studies used lonely or in combination with radial EUS and that might generate heterogeneous. However, our meta-regression for EUS type did not show statistical significance.

EUS and CT technology for diagnosing vascular invasion have changed over the decade. We grouped included studies into three time periods to standardize the technology and vascular invasion criteria. An interesting finding was that in more recently published studies (2000–2005 and 2006–2013), the diagnostic performance of EUS was significantly worse than in earlier publications. One reason for decreasing the accuracy over time might be due to absence of sufficient studies in the earlier time periods. Another possible explanation was that technology and study design have improved over time requesting more critical evaluation, which possibly lead to a lower diagnostic performance. The sensitivity and specificity of CT were up to 74 and 90 % in the recent 6 years (2000–2005). The explanation should be more advanced technology, and diagnostic criteria applied in clinical trial (Klauss et al. 2008; Karmazanovsky et al. 2005; Lee et al. 2010).

The strength of this meta-analysis was that we collected the data in assessing venous invasion alone (portal vein with confluence, superior mesenteric vein (SMV), and splenic vein). It was noteworthy that the criterion for arterial invasion of EUS was different from venous invasion (Snady et al. 1994; Mertz et al. 2000), which might lead to a lower accuracy. Although arterial invasion was defined endoscopic ultrasound by alteration of vessel course and caliber at the site of the cancer (Mertz et al. 2000), this course often difficult to define. EUS has been suggested to be more sensitive in detecting venous invasion. A prospective study by Midwinter et al. found that the sensitivity and specificity of EUS in detecting venous invasion were 81 and 86 %, in contrast to 17 and 67 % for arterial invasion. CT was considered more sensitive in detecting arterial invasion on preoperative evaluation in pancreatic cancer (Rosch et al. 2000; Midwinter et al. 1999; Kala et al. 2007). However, only scarce comparative data were reported especially for arterial invasion, and we could not perform a subgroup comparison between EUS and CT scan in assessing venous versus arterial invasion.

There were several limitations to our study. First, the moderate sample size for the statistical analysis was limited. In practice, larger numbers of participants were difficult to obtain because the incidence of pancreatic cancer was 10 per 100000, and only about 10 % of pancreatic cancer could undergo curative surgery. Second, we were not able to analyze the inherent heterogeneity in the study design. The optimal protocol for performing EUS or CT had not been defined. In this meta-analysis, four different EUS criteria have been used in studies to define vascular invasion. Third, several studies included in our meta-analysis contained ampullary tumors and other subtypes, which could lead to heterogeneous data. Differentiation of ampullary and duodenal papillary malignancy with pancreatic cancer might be difficult in many studies, because of both forms of malignancy could infiltrate into pancreas.

In conclusion, based on the currently available English language literature, our meta-analysis suggested that EUS performed better than CT in differentiating vascular invasion preoperative on pancreatic cancer. However, to minimize the population bias, our head-to-head comparison showed that EUS appeared to be more accurate in the diagnosis of vascular invasion (AUC 0.9379 vs. 0.8972). High-quality trails were still needed to be conducted. With the advanced spiral CT applied in clinical trial, the sensitivity and specificity of CT were promoted obviously. In addition, EUS had other benefit over CT, such as EUS has the ability to obtain tissue samples for pathological diagnosis. For this reason, we recommend that all patients with pancreatic cancer be evaluated by EUS preferential for vascular invasion.