Introduction

Hepatocellular carcinoma (HCC) is the most common primary liver cancer in patients with chronic liver disease (CLD), and is typically thought to be caused by viral infection (hepatitis B or C) and alcoholism. However, such patients are also at significantly increased risk of intrahepatic mass-forming cholangiocarcinoma (IMCC) [1–3]. In addition, combined hepatocellular-cholangiocarcinoma (CHC) has many demographic and clinical similarities to IMCC, and also shares etiological risk factors with HCC [4, 5]. Various therapeutic methods, including liver transplantation, surgical resection, radiofrequency ablation (RFA), and transarterial chemoembolization (TACE), can be applied for the treatment of HCC [6]. However, for IMCC and CHC, surgical resection is the most effective and is the only therapy associated with prolonged disease-free survival, whereas liver transplantation is generally contraindicated because of the high rate of recurrence [7–11].

Advances in MR imaging technology and the development of tissue-specific contrast material such as gadoxetic acid have made it possible to differentiate most HCCs from IMCCs and CHCs based on their typical signal intensities and enhancement patterns [12–19]. A diagnosis of HCC is currently based primarily on imaging findings according to the guidelines established by the American Association for the Study of Liver Diseases (AASLD) [6]. However, HCCs can show atypical enhancement patterns on gadoxetic acid-enhanced MR imaging [17, 20, 21], and small IMCCs can show arterial hypervascularity with delayed washout similar to that in HCC [18, 19]. In addition, imaging features of CHC may overlap with those of classic HCC or IMCC [10, 18, 22]. Although percutaneous liver biopsy prior to treatment is typically considered for differentiation of IMCC and CHC from HCC, seeding to the peritoneum or along the needle tract has occasionally been reported [23, 24]. Therefore, in high-risk CLD patients, it is important to identify predictors for discriminating IMCC and CHC from HCC in primary liver cancer using gadoxetic acid-enhanced MRI.

For this purpose, we employ classification tree analysis (CTA). CTA is a machine-learning method of predicting the class levels (e.g., IMCC and CHC) of an outcome variable based on several predictors. The advantages of CTA include (1) its robustness to normalization, scaling, and outliers of data, (2) the lack of a need for linearity assumptions in the data, and (3) its ability to visually and explicitly represent the prediction process [25].

Accordingly, the purpose of this study was to identify the predictors for the discrimination of IMCC and CHC from HCC for primary liver cancers on gadoxetic acid-enhanced MRI among a group of high-risk CLD patients through the use of CTA.

Materials and methods

Study population

This retrospective study was approved by our institutional review board, and the requirement for informed consent was waived. Using our institutional surgical pathologic data, we retrospectively searched cases of IMCC and CHC between January 2012 and January 2015, and HCC between January 2014 and January 2015. Those meeting the following inclusion criteria were then selected for this study: (a) patients who met the Milan criteria [26], (b) who were diagnosed with chronic liver disease according to pathologic or clinical criteria [27], (c) who had no history of previous treatment for hepatic tumour, and (d) who underwent gadoxetic acid-enhanced liver MR imaging according to our standard protocol within 1 month prior to surgery (Fig. 1). The median length of time between MR imaging and surgery was 17 days (range, 2–26 days). A total of 152 patients with 155 tumours were included in this study. Among these, 64 patients with 40 IMCCs and 24 CHCs were defined as the cholangio-combined group, and 90 patients with 91 HCCs were defined as the HCC group. Three patients had two tumours each (two HCCs, one IMCC and one HCC, and one CHC and one HCC). The clinical and demographic data of the study groups are summarized on Table 1.

Fig. 1
figure 1

Flowchart of the inclusion process for the study group. IMCC intrahepatic mass-forming cholangiocarcinoma, CHC combined hepatocellular-cholangiocarcinoma, HCC hepatocellular carcinoma. *Milan criteria = a solitary tumour with a diameter less than 5 cm, or up to three hepatic nodules, each smaller than 3 cm, with no evidence of vascular invasion or extrahepatic metastasis according to preoperative radiologic findings

Full size image
Table 1 Clinical and demographic data
Full size table

The preoperative serum levels of α-fetoprotein (AFP) and protein induced by vitamin K absence or antagonist-II (PIVKA-II) within 1 month prior to surgery were recorded. At our institution, elevated AFP and PIVKA-II levels are defined as serum levels higher than 8.1 ng/mL and 40 mAU/mL, respectively. To determine patient eligibility for enrolment for this study, a review of pathologic reports, clinical data, and MR images was performed by one abdominal radiologist (H.J.P., with 9 years of experience in abdominal MRI interpretation).

Image acquisition

A 3-T whole-body MRI system (Intera Achieva 3.0 T; Philips Healthcare, Best, Netherlands) with a 32-channel phased-array (Torso/Cardiac, Philips Healthcare) receiver coil was used in this study. Liver images for all patients were acquired in the axial plane both before and after administration of gadoxetic acid at a dose of 0.1 mL/kg (0.025 mmol/kg). The contrast agent was administered intravenously at a rate of 2 mL/s using a power injector, followed by a 20-mL saline flush. The MRI protocol is summarized in Table 2. Diffusion-weighted imaging (DWI) was performed before the administration of gadoxetic acid using a respiratory-triggered single-shot echo planar imaging with b-values of 0, 100, and 800 s/mm2. A spectral attenuated inversion recovery technique was used for fat suppression on DWI. The apparent diffusion coefficient was calculated using a monoexponential function with b values of 100 and 800 s/mm2 in order to minimize perfusion effects. For gadoxetic acid-enhanced MRI, unenhanced, enhanced arterial phase (20–35 s), portal phase (60 s), 3-min delayed phase, and 10-min and 20-min delayed hepatobiliary phase images (HBPI) were obtained using a T1-weighted 3D turbo field-echo sequence (THRIVE [T1 high-resolution isotropic volume examination]; Philips Healthcare) with spectral attenuated inversion recovery fat suppression. The timing for the arterial phase imaging was determined using MR fluoroscopic bolus detection.

Table 2 MRI sequences and parameters
Full size table

Image analysis

All MR images were retrospectively and independently reviewed on a commercial workstation with a 2,000 × 2,000 PACS monitor (Centricity; GE Healthcare) by two abdominal radiologists (T.W.K. and K.D.S., with 10 and 7 years of experience in abdominal MRI interpretation, respectively) in a blinded manner. These reviewers were unaware of the pathological tumour diagnoses. In cases of disagreement, a third observer (S.H.K., with 16 years of experience in abdominal MRI) was asked for an opinion, and a majority decision was reached which was used for data analysis.

The MR imaging features that had been evaluated for investigation of IMCC, CHC, and HCC in previous studies [4, 9, 11, 13, 14, 16–19, 28, 29] are summarized in Table 3 (Figs. 2, 3, and 4). For the exclusion of old haemorrhage in hypointense areas on T2WI, we correlated the area with other sequence images. In addition, arterial rim enhancement persisting throughout portal venous and 3-min delayed phases was not regarded as capsule appearance. We evaluated dynamic enhancement patterns in the solid portions of tumours except for areas of haemorrhage and necrosis. A type I enhancement pattern was defined for a tumour showing arterial phase hyper-enhancement and washout appearance on portal venous or delayed phases. A type II enhancement pattern was defined for a tumour showing progressive enhancement or an unclassifiable enhancement pattern. The enhancement pattern referred to as “washout” was defined for lesions showing temporal reduction in enhancement relative to the liver from an earlier to a later phase resulting in portal venous phase or delayed phase hypo-enhancement [29]. Progressive enhancement was defined as persistent or gradual enhancement of a tumour on portal venous and 3-min delayed phase images compared with arterial phase images. Tumours with combined type I and type II enhancement patterns were regarded as having a type II enhancement pattern.

Table 3 Assessed MR imaging features and their definitions according to the favoured tumour group
Full size table
Fig. 2
figure 2

Hepatocellular carcinoma (HCC, top, a–c) in left lateral liver with elevated PIVKA-II and combined hepatocellular-cholangiocarcinoma with HCC predominant type and normal tumour markers (CHC, bottom, d–f) in segment IV. HCC (white arrows) shows globular shape, aggregated tiny hyperintense foci on T2-weighted image (a) and type I enhancement pattern (arterial hyper-enhancement [b] and washout appearance on portal venous phase image [c]). CHC (black arrows) shows non-globular shape on T2-weighted image (d) and type I enhancement pattern on arterial (e) and 3-min delayed phase (f) images

Full size image
Fig. 3
figure 3

A 57-year-old man with hepatocellular carcinoma in right lobe and normal tumour markers. The tumour (arrows) shows signal loss on out-of-phase imaging (a) compared with signal intensity on in-phase imaging (b), suggesting intralesional fat. Persistent hyperintense area (curved arrows) in the images is suggestive of hemorrhagic foci. Although the tumour shows arterial rim enhancement (c) and type II enhancement pattern on 3-min delayed-phase image (d), additional findings suggest HCC as globular shape, hyperintense area on T1-weighted image, intralesional fat, and hyperintensity on 20-min hepatobiliary phase image (e)

Full size image
Fig. 4
figure 4

A 63-year-old man with cholangiocarcinoma in segment IV and normal tumour markers. The tumour (arrows) shows non-globular shape, arterial rim enhancement and type II enhancement pattern on arterial (a) and 3-min delayed phase images (b). Target sign on DWI (b = 800 s/mm2) (c) and 10-min hepatobiliary phase image (d) are also identified. Asterisk (*) indicates gallbladder

Full size image

Statistical analysis

MR imaging features and levels of AFP and PIVKA-II were analyzed for comparison between the cholangio-combined and HCC groups using Wilcoxon rank-sum and Pearson chi-square tests. To ascertain the values of tumour markers and parameters for differentiating between the two groups, we used univariate logistic regression. Each logistic regression model provides P values and 95 % CI for odds ratios (OR).

For explicit identification of high-risk factors in the cholangio-combined group, a CTA was performed. The prediction model is obtained through a process described as follows. First, a single variable is found which splits the data into two subsets and which also minimizes impurities across the two subsets. In CTA, a set is pure if all the data in the set have a common outcome variable class level. The data are separated, and the process is then applied recursively to each derived subset. This recursive partitioning continues until the subsets reach a minimum size or no further improvement is expected. The partitioning can thus be represented graphically as a decision tree. After verification of the prediction tree model based on fivefold cross-validation (CV), we present a final model in which all the data is incorporated. We calculate sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV), and diagnostic accuracy to evaluate our developed model [25].

Interobserver agreement regarding the MR imaging features of the lesions was evaluated by kappa (κ) statistics. A kappa value less than 0.20 indicated poor agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, good agreement; and greater than 0.81, excellent agreement. Classification tree analysis was performed using the package “rpart” in R version 3.0.2 (Vienna, Austria; http://www.R-project.org/). All other statistical analysis was performed using SAS version 9.4 software (SAS Institute, Inc., Cary, NC, USA). All P values were two-sided, and P < 0.05 was considered statistically significant.

Results

The levels of AFP and PIVKA-II were significantly different between the cholangio-combined group (mean, 9.46 ng/mL, 28.08 mAU/mL, respectively) and HCC group (mean, 490.8 ng/mL, 759.3 mAU/mL, respectively; P = 0.0109) (Table 1). However, the levels of AFP and PIVKA-II were elevated in 17 (26.6 %) and 5 (7.8 %) tumours, respectively, in the cholangio-combined group.

The MR imaging features for both groups are summarized in Table 4. Although there were significant differences between the two groups for all assessed parameters on MRI, most parameters were not perfect predictors for differentiating between the cholangio-combined and HCC groups. Intralesional fat was the only parameter shown exclusively in the HCC group, occurring in 24 (26.4 %) tumours. A type I enhancement pattern, which is considered typical of HCC, was shown in 12 (18.8 %) tumours of the cholangio-combined group, whereas 19 (20.9 %) tumours in the HCC group showed type II enhancement patterns (Figs. 2 and 3). Most of the assessed parameters on MRI showed good or excellent interobserver agreement, but interobserver agreement for enhancement pattern and the presence of fibrosis on T2WI was fair (κ = 0.599) and poor (κ = 0.271), respectively (Table 4). The OR of each imaging parameter for the differentiation of the two groups is also summarized in Table 4. Arterial rim enhancement, at 18.585, showed the highest OR for differentiating the cholangio-combined group from the HCC group.

Table 4 Qualitative and logistic regression analyses
Full size table

In the CTA prediction model, arterial rim enhancement was the first splitting predictor for assessing the probability of a tumour being classified in the cholangio-combined group (Figs. 4 and 5). For tumours with arterial rim enhancement, the following predictors were no intralesional fat, non-globular shape of the tumour, target sign on DWI, no elevation of AFP level, and no hyperintensity on T1WI (Fig. 3). Of 43 tumours with arterial rim enhancement, no intralesional fat, and non-globular shape, 41 (95.3 %) were either IMCC (n = 29) or CHCs (n = 12). For tumours with no arterial rim enhancement, the following predictors were capsule appearance, elevated AFP level, no target sign on DWI, no liver surface retraction, elevated PIVKA-II level, and no biliary dilatation. The classification tree model established in this study has sensitivity of 84.4 %, specificity of 97.8 %, PPV of 96.4 %, NPV of 89.9 % and diagnostic accuracy of 92.3 % for discriminating the cholangio-combined group from the HCC group. In the HCC group, only two (3.2 %) tumours were classified in a high probability of cholangio-combined subgroup.

Fig. 5
figure 5

Subgroups of the tumours identified through classification tree analysis and their risk of development for each group. Data are number of lesions. Numbers in parentheses are percentages. Group A = cholangio-combined group; Group B = hepatocellular carcinoma group. Boxes with bold double-line depict the subgroup of tumours with high probability of classification in the cholangio-combined group. Boxes with dotted line depict the subgroup of tumours with high probability of classification in the hepatocellular carcinoma group. E enhancement

Full size image

Discussion

In this study, MR imaging features that had been evaluated for investigation of IMCC, CHC, and HCC in previous studies and tumour markers for HCC were significantly different between the IMCC and CHC group and the HCC group, although none but intralesional fat was a perfect predictor for the differentiation of IMCCs and CHCs from HCCs. Therefore, we sought to identify valuable combinations of predictors for differentiating IMCCs and CHCs from HCCs using a particular statistical methodology: CTA. We were able to establish a predictive algorithm with sensitivity of 84.4 %, specificity of 97.8 %, and diagnostic accuracy of 92.3 % for discriminating IMCCs and CHCs from HCCs.

CTA is a non-linear, non-parametric alternative to linear models for classification problems, in which classification trees are built for predicting class labels of outcome variables. In these tree structures, each interior node corresponds to one of the predictor variables. Leaves represent class labels of the outcome variable, and branches represent combinations of features that lead to those class labels. CTA offers a means for simple and explicit interpretation of results and judgment processes. In general, because it provides dichotomous cut points of predictor variables, CTA can provide class labels for outcome variables that are routinely required in clinical assessments. It also facilitates the identification and interpretation of complex interactions, whereas other multivariate analysis methods can handle only those interactions predetermined by the analyst. Several similar models have been revealed to be efficient tools for discriminating benign from malignant tumours or for identifying groups at high risk of a disease [30–32].

In this study, the typical enhancement pattern of HCC was not used as a predictor in CTA for assessing the probability of a tumour belonging to the cholangio-combined group. Although the typical HCC enhancement pattern is a valuable diagnostic criterion according to AASLD guidelines [6], atypical enhancements patterns on gadoxetic acid-enhanced MRI have been reported in 17.7–28.3 % of cases in previous studies [20, 21]. In the present study, 19 (20.9 %) HCCs showed atypical enhancement patterns. Differentiation of IMCCs and CHCs from atypical HCCs, therefore, can be challenging. The target sign on DWI was reported as a valuable imaging feature for distinguishing IMCCs from HCCs in a previous study [19], and it was used as a predictor for assessing the probability of classifying a tumour in the cholangio-combined group on CTA. In the CTA analysis of our study, 89 (97.8 %) tumours in the HCC group were classified as having a high probability for HCC, regardless of the tumour enhancement pattern.

In the present study, arterial rim enhancement, which showed the highest OR of 18.585 for differentiation between the cholangio-combined and HCC groups, was the first splitting predictor for assessing the probability of a tumour being classified as cholangio-combined on CTA. In a previous study by Jeong et al. [33], arterial rim enhancement was observed in 84–89 % of IMCCs, and was attributed to a greater density of viable tumour cells in the periphery with central fibrous stoma. Hwang et al. [34] reported that all 20 IMCCs and CHCs in their study showed strong or weak arterial rim enhancement. In addition, although 11 (12.1 %) tumours in the HCC group also showed arterial rim enhancement, nine (81.8 %) of those 11 were correctly classified as having high probability for HCC based on the presence of intralesional fat, globular shape, absence of target sign on DWI, elevation of AFP level, and presence of hyperintensity on T1WI. Therefore, when arterial rim enhancement is observed in a tumour in patients with CLD, the first differential diagnosis considered should be IMCC or CHC, although it is necessary to identify additional findings suggestive of HCC, as mentioned previously. In recent years, arterial rim enhancement in HCCs has been considered a predictor of rapid progression, early recurrence, poor cellular differentiation, and worse tumour prognosis [17, 35, 36]. Therefore, surgical resection may be preferred over other treatment modalities, including TACE and RFA, for treatment of tumours with arterial rim enhancement in patients with CLD, although larger, more sophisticated studies would be needed.

In this study, none of the tumours in the cholangio-combined group showed intralesional fat. The presence of intralesional fat has been considered an ancillary feature that favours the diagnosis of HCC and a marker of carcinogenesis from dysplasia to HCC on CT or MRI [29, 37, 38]. This feature is attributable to a decrease in intratumoral arteries without alteration of intratumoral portal tracts [39]. Aasayama et al. reported the absence of fat as a key characteristic finding for IMCC, which may help its differentiation from poorly differentiated HCC [40].

This study has several limitations. First, selection bias may exist because of the retrospective study design, which included only patients with IMCC, CHC, or HCC who had undergone gadoxetic acid-enhanced MRI and hepatic surgery. Second, also because of the retrospective nature of the study, an exact correlation between the radiologic features and overall pathological composition was not performed. Third, as tumours included in this study met the Milan criteria, they were less than 5 cm in diameter. Therefore, the question remains whether our results would be useful in differentiating cholangio-combined from HCC groups for tumours greater than 5.0 cm in diameter. Fourth, the diagnostic performance of the CTA model in this study may be overestimated, as the performance results were obtained in the same study population as that for which the CTA model was developed. Therefore, validation of the model in a different study population or internal cross-validation would be needed to demonstrate whether the CTA model can be generalized to other population. In the present study, the performance of our CTA model was assessed based on internal validation (i.e., fivefold CV), and therefore our model may be over-fitted.

In conclusion, we established a simple classification tree model for differentiating IMCCs and CHCs from HCCs in primary liver cancers among a group of high-risk CLD patients. This classification tree model may be useful for guiding the diagnosis of primary liver cancers in patients with CLD.