A systematic literature search using PubMed (https://pubmed.ncbi.nlm.nih.gov), Embase (https://www.embase.com), Web of Science (https://apps.webofknowledge.com) and Cochrane Library databases (https://www.cochranelibrary.com) were performed independently by three radiologists to identify articles published prior to June 2019, using the key words ‘hepatocellular carcinoma’, ‘liver cancer’, ‘liver cell carcinoma’, ‘magnetic resonance imaging’, ‘diffusion magnetic resonance imaging’, ‘diffusion MRI’ and ‘diffusion weighted MRI’. Related citations in eligible articles were also assessed for inclusion.

The three radiologists reviewed all 2447 abstracts after duplication removal and subsequently the full text of the 119 articles was obtained if the following inclusion criteria was fulfilled: i) Included the diagnostic accuracy of MRI with DWI for HCC; ii) constituted original research rather than a meta-analysis, a review article, case report or case series; iii) published in English; and iv) results are from humans and not animals. For the studies in which the full text was reviewed, the following inclusion criteria was used: i) Included original data regarding the detection of small HCC lesions, ≤2 cm; ii) included both contrast-enhanced MRI and DWI; iii) included sufficient data, with >20 patients to calculate true positive (TP), false positive (FP), false negative (FN) and true negative (TN) for constructing a 2×2 contingency table; and iv) patients diagnosed with hepatic lesion using pathological analysis (surgical resection, explant and/or biopsy) or imaging from follow-up according to the guidelines for standardization of liver imaging, diagnosis, classification and reporting of hepatocellular carcinoma (3). In addition, articles from the same institution, which included an overlap period of patient recruitment were considered to have an overlapping population. In these cases, the study, which had the larger number of small HCCs cases, was included. If there were disagreements between the three investigators, the consensus amongst the three radiologists was used to resolve the disagreement. Disagreements were resolved following discussions between the three investigators, until at least two of the investigators reached the same conclusion. Attempts were made to contact the authors of the article only if data for the 2×2 contingency table was not fulfilled from the inclusion criteria (authors of two articles had been contacted for this study). A total of 109 studies were excluded according to the following exclusion criteria: i) They were not relevant to the present meta-analysis if they fit one of the followings conditions: Cancer type includes malignant cancer other than HCC, such as cholangiocarcinoma, hepatoepidermoid carcinoma and metastatic cancer; diagnosis of HCC using a combination of multiple imaging modalities; size of HCC lesions >2 cm; ii) the size of the HCCs was not specified; iii) they evaluated previously treated HCCs; iv) the specificity was not evaluated; v) there was a lack of sufficient data to construct a 2×2 contingency table; and vi) there was study population overlap. A total of 10 studies were included for analysis. In addition, the reference list of these 10 studies was reviewed. Once any of the studies fulfilled the inclusion criteria but not the aforementioned exclusion, they were not included for analysis. No studies were excluded in the process.

A total of two investigators reviewed the included studies and extracted the relevant details for the meta-analysis. The study characteristics extracted included the authors of the study, year of publication, country of origin, number of overall patients, overall size of HCC, cause of liver cirrhosis, study design (prospective or retrospective image interpretation), study period, b value of DWI, MRI field strength, number of HCCs which were ≤2 cm, number of benign lesions, type of benign hepatic lesions, reference standard and number of readers. The number of readers is important as diagnosing HCC lesions by multiple readers or radiologists increases the accuracy of the diagnosis, which improves the reliability of the studies.

Data for the diagnostic value of DWI combined with contrast-enhanced MRI for small HCC lesions were extracted to construct a 2×2 contingency table. If the sensitivity and specificity were reported by multiple radiologists, the average sensitivity and specificity scores were reported to avoid under- or overestimation of the diagnostic accuracy, which occurred in 1 out of 10 of the included studies. In addition, raw data for the diagnostic value of contrast-enhanced MRI was extracted if available for the construction of a 2×2 contingency table, which was available in 8 out of the 10 included studies. For studies with DWI, the information of whether a preset apparent diffusion coefficient (ADC) cutoff value to diagnose HCC was also extracted for analysis, which was available in 1 out of 10 included studies). All the data were analyzed using Stata version 14.0 (https://www.stata.com).

The quality of the included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) (11). Exclusions of patients with lesion smaller (<1 cm) was considered inappropriate since this increased the selection bias. Any reference standard to diagnose small hepatic lesions (≤2 cm), other than pathological analysis (biopsy, surgical resection and explant), for example, imaging follow-up, was considered unlikely to lead to correct classification of the target condition and may introduce bias. Diagnosis using reference standard based on imaging follow-up or subsequent transcatheter arterial chemoembolization were considered to include the knowledge of the results of the index test. In addition, the risk of bias for reference standard results based on biopsy or resection were considered unclear due to the lack of information provided to the pathologist at the time of assessment. An interval of >90 days between MRI scan and reference standard examination was considered inappropriate since during such a long interval, new tumorous growth adjacent to the targeted mass identified by MRI scan or reference standard examination may happen; the targeted mass may become larger over 2 cm; patients may receive treatment that may change the size and the cell composition of the targeted mass. All the aforementioned criteria were used by the two evaluators to specify the QUADAS-2, which was generally developed for quality assessment for all meta-analysis, to assess the quality of the included studies for the present analysis.

All the data were analyzed using Stata software (version 14.0; College Station, TX, USA). The sensitivity, specificity and 95% confidence intervals (CIs) for the diagnosis of small HCC lesions using contrast-enhanced MRI with DWI were calculated using the bivariate random effects model (12), which were demonstrated via forest plots. The assessment for the sensitivity and specificity was performed on a per-lesion basis for all the included studies. The summary receiver operating characteristic curves (SROCs) were constructed and the area under the SROCs (AUC) of the conventional MRI with DWI and conventional MRI alone were calculated to determine the diagnostic performances. χ² test (P<0.05 indicating significant heterogeneity) and I² was used to determine the heterogeneity. A random effects model was used if I² >50%; otherwise, a fixed effects model was used. Univariate meta-regression analysis was performed according to MRI field strength (1.5 T vs. 3.0 T), country of origin (Asia vs. non-Asia), study design (prospective vs. retrospective) and whether hepatobiliary phase was used in the diagnosis of small HCC. In Asia countries, Hepatitis B virus infection is the leading cause of cirrhosis which results in HCC (13). In non-Asia country, the etiology of HCC varies (13). The difference in the etiology of HCC may be the cause of the heterogeneity. This was the reason that country of origin was divided into Asia vs. non-Asia in the present study to assess the potential cause of heterogeneity. In addition, the diagnostic value of DWI combined with contrast-enhanced MRI was assessed by comparing the diagnostic performance of DWI with contrast-enhanced MRI to contrast-enhanced MRI only. The calculation of TP, FP, TN and FN for the contrast-enhanced MRI was only available for 8 of the 10 included studies (7,8,14–19). Multilevel mixed-effects logistic regression analysis was used to compare the summary paired sensitivity/specificity data with a significance level of P<0.05. Publication bias was assessed using a Deeks' funnel plot.

A flow chart following the Preferred Reporting Items for Systematic Reviews and Meta-analysis principles was used to demonstrate the selection procedure (Fig. 1). A total of 3,567 articles were initially identified. There was a total of 2,447 articles remaining following the removal of duplicates and a further 2,328 articles were excluded, following screening of the abstract. Amongst the remaining 119 studies, a total of 10 studies were included in the meta-analysis using the inclusion criteria (7,8,14–21).

The summarized characteristics and the diagnostic performance of DWI combined with contrast-enhanced MRI for the included 10 studies are shown in Tables I and II, respectively. A total of 837 small HCCs with a diameter ≤2 cm and 545 benign liver lesions, with a diameter ≤2 cm was included in the meta-analysis. The TP, FP, FN and TN were all calculated on a per-lesion basis. Of the included studies, 6 originated from Asia, 3 from Europe and one from Egypt. In addition, seven of the studies were retrospective, and three were prospective. The reference standard for the diagnosis of HCC included pathological analysis (surgical resection, explant and/or biopsy) and imaging from follow-up. MR imaging field strength was all ≥1.5 T.

Fig. 2 demonstrates the overall evaluation for the quality of the included studies using QUADAS-2. The quality of the index test was high (90%, 9/10 studies); however, patient selection had a low score (70%, 7/10 studies), which could be due to a lack of avoidance of a case-control design or the inappropriate exclusions during patient selection. This also increased concerns regarding the applicability of patient selection. For all the 10 included studies, some of them used pathological finding as the only reference standard to diagnose HCC. For the others, imaging follow up was used as a reference standard for patients when pathology analysis was not available. A low score was found for the reference standard (60%, 6 of 10 studies) due to a lack of using pathological analysis as a reference standard. However, these studies used imaging follow-up as one of the reference standards for those without pathological confirmation, which has been shown to be effective for the diagnosis of HCC (22). Therefore, the concerns of bias for applicability of reference standards was low for studies using imaging follow up as one of the reference standards for patients when pathology was not used. The risk of bias for flow and timing was high for 1 study since the interval between MRI scan and the pathological analysis exceeded 90 days for some of the patients, and was unclear for 2 studies for the lack of information regarding the time interval between MRI scan and the reference standard. The Deeks' funnel plot (Fig. 3) suggested that there was no significant publication bias (P>0.05).

The 10 included studies demonstrated significant heterogeneity with P<0.001 using χ² test. The heterogeneity for the sensitivity (I² of 85.7) was higher compared with that for specificity (I² of 78.11). In addition, there was no threshold effect found (correlation, −0.65; proportion of heterogeneity due to threshold effect, 0.42).

Fig. 4 demonstrates the forest plots of sensitivity and specificity for DWI combined with conventional MRI for the diagnosis of small HCC lesions, with a diameter ≤2 cm. The pooled sensitivity and specificity were 0.88 (95% CI, 0.80–0.93) and 0.90 (95% CI, 0.81–0.95), respectively. The positive and negative likelihood ratio was 8.4 (95% CI, 4.6–15.3) and 0.13 (95% CI, 0.08–0.22), respectively. Fig. 5 shows the summary ROC curve with an AUC of 0.95.

The results of the univariate meta-regression analysis are shown in Table III. Sensitivity was significantly higher for studies not using hepatobiliary phase compared with those using hepatobiliary phase (P<0.001). Specificity was significantly higher for studies using a 3 T magnetic field compared with those using 1.5 T magnetic field (P=0.03). There were no significant differences in either the sensitivity or in specificity for the remaining study characteristics (all P>0.05).

The comparisons in the diagnostic performance of the different combinations of MRI protocols in the diagnosis of small HCC lesions are shown in Table IV. The sensitivity of DWI with contrast-enhanced MRI was significantly higher compared with that in contrast-enhanced MRI alone (0.89 vs. 0.78; P=0.01). However, there was no significant difference for the specificity between DWI with contrast-enhanced MRI and conventional MRI alone (P=0.603).