From 2006 to 2012, we searched electronic medical record in our hospital consecutively to identify eligible patients who met following criteria: i) adult patient (≥ 18 years), ii) pathologically confirmed perihilar cholangiocarcinoma, iii) who underwent curative-intent surgery and iv) pre-operative multiphase abdominal CT. Exclusion criteria were as follows: history of other malignancy including perihilar or elsewhere, no recent CT within 4 weeks before surgery, or suboptimal CT image quality. According to the criteria, a total of 139 patients (82 men and 57 women; mean age, 66.9 years, age range, 42-75 years) were selected for this study (Fig. 1). All 139 patients underwent curative-intent surgery by one of two pancreatobiliary surgeons with 27 and 18 years of experience in pancreatobiliary surgery in the same institute. The extent of surgery were preoperatively discussed with board certified surgeons, radiologists and internists in a multidisciplinary team meeting The mean interval (± standard deviation) between last CT examination and surgery was 17.5 days ± 11.0 (range, 0 to 28 days). Patients underwent extended right hepatectomy (n = 48), extended right hepatectomy with pancreaticoduodenectomy (n = 2), extended left hepatectomy (n = 26), extended left hepatectomy with pancreaticoduodenectomy (n = 1), hilar resection (n = 56), hilar resection with pancreaticoduodenectomy (n = 2), bypass or open and closure (n = 4). Detailed patient demographics are summarized in Table 1.

All patients underwent four-phase multidetector CT (MDCT) (n=96) consisting of pre-contrast, early arterial, late arterial and venous phase, or three-phase MDCT (n=43) consisting of pre-contrast, arterial and venous phase before curative-intent surgery. Owing to retrospective nature of this study, various CT scanner were used. Detail parameters of each CT scanner were noted in Appendix 1. Contrast media was injected at a rate of 2.0 to 3.0 mL/sec using a power injector with a total volume of 1.5 mL per kilogram of the patient body weight followed by flushing with 20 mL normal saline. After contrast media injection, an automatic bolus tracking technique was used for image acquisition. The trigger threshold was 100 HU at the abdominal aorta. Early arterial phase was obtained 6 seconds after the trigger threshold and pancreatic phase was obtained 5 to 9 seconds after the early arterial phase. Venous phase was obtained 40 to 50 seconds after early arterial phase.

Two board-certified radiologists (I.J. and W.C, with 12- and 9-years of experience in abdominal imaging) independently reviewed all pre-operative MDCT images. They were blinded to patients' histological or operative results, but were aware that this study population had PHC. Reviewers were requested to determine the likelihood of tumor involvement of bile duct (BD) in longitudinal and transverse plane. For longitudinal plane assessment, the likelihood of tumor involvement from intrapancreatic common bile duct (CBD) to bilateral second order branches of intrahepatic bile duct (IHD) using 5-point scale as follows: score 1, definitely no, no significant image findings; 2, probably no, presenting only short stricture with smooth transition; 3, indeterminate, one positive image finding suggestive of malignancy; 4, probably yes, two positive image findings suggestive of malignancy; 5, definitely yes, more than three positive image suggestive of malignancy. The positive image findings suggestive of malignancy are long-segmental involvement (> 12 mm), a thickened (> 1.5 mm) wall, luminal irregularity or asymmetry, and incremental enhancement may favor a malignant stricture [21, 22]. For evaluating tumor involvement in the transverse plane to BD, following 5-point scale was used: score 1, definitely no, no significant image findings; 2, probably no, minimally increased attenuation in peritumoral fat; 3, indeterminate, increased attenuation with ground-glass like pattern in peritumoral fat; 4, probably yes, increased attenuation with reticular pattern in peritumoral fat; 5, definitely yes, obviously increased attenuation as much as BD wall [23]. The cutoff for tumor involvement was determined by cutoff values from the receiver operating curve analysis. Mid-CBD was defined as CBD below cystic duct or more than 2 cm away from 1st bifurcation of CBD.

Also, the relation between tumor and vessels (proper, left and right hepatic artery [PHA, LHA, and RHA]; main, left and right portal vein [MPV, LPV, RPV]) was evaluated using three categories: no relation, abutment (180 degrees or less of a vessel's circumference), encasement (more than 180 degrees of tumor involvement of a vessel's circumference). In artery evaluation, abutment or encasement classified as tumor involvement. While vein evaluation, encasement, change caliber of vessels in the region of contact, or irregularity of vessel margin were assessed as tumor involvement [24]. Furthermore, the anatomy type of BDs were assessed according to the five anatomy types reported by Huang et al. [25] namely, normal, trifurcation, right posterior BD draining to the left hepatic duct, right posterior BD draining to the common BD, and right posterior BD draining to the cystic duct. The conventional BD anatomy was defined as the right posterior BD draining to the right hepatic duct, with the right and left hepatic ducts joining the common hepatic duct. In addition, Bismuth type [26], and LN involvement (no significant LN enlargement, indeterminate LN, highly suspected metastatic LN) were also assessed. When LNs larger than 10 mm in short diameter, central necrosis or hyper-attenuating than liver in portal venous phase were considered highly suspected metastatic LN [14, 27]. Indeterminate LN was defined when only one positive findings of metastatic LN was noted.

The pre-operative level of total bilirubin, albumin, alkaline phosphatase, carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA 19-9) have been routinely checked for pre-operative planning of perihilar cancer in our hospital. When the patients received a biliary intervention to decompress the bile duct before surgery (n = 25, 18.0%), the laboratory results after biliary intervention were used for analysis.

For histologic analysis, all surgical specimens were reviewed by an experienced pathologist (K.L. with more than 17 years' experience in pancreaticobiliary pathology). Information of tumor location, tumor size, depth of invasion, lymph node involvement, angiolymphatic, venous, or perineural invasion, and surgical margin were routinely reported in the pathologic reports at our hospital. After curative-intent surgery, microscopically (R1) or marcroscopically (R2) residual tumor was noted in 68 patients, while 71 patients did not (R0). The survival data were obtained from the official death reports issued by the government.

To determine predictive factor for margin positive resection (R1 or R2) on pre-operative CT, a chi-square test was used for univariable analysis (likelihood of BD involvement from intrapancreatic CBD to bilateral second-order branches, peritumoral fat infiltration, the relation between tumor and vessels, and LN involvement). The parameters that proved to be significant on the univariable analysis (tail p-values less than 0.1) were then subsequently tested with the multivariable using logistic regression. The survival rate during the follow-up period was reported with a 95% confidence interval using the Kaplan-Meier method. After univariable analysis of each parameter (pre-operative image findings, laboratory results, and extent of surgery) using a log-rank test, the parameters that proved to be significant (tail p-values less than 0.1) were then subsequently tested with the multivariable analysis using Cox proportional hazard model. Interobserver agreement was assessed using weighted κ statistics for noncontinuous scale and intraclass correlation coefficient (ICC) for continuous scale. The strength of agreement was evaluated as follows: a κ value and ICC value of less than 0.20 indicated poor agreement; 0.21-0.40, fair agreement; 0.41-0.60, moderated agreement; 0.61-0.80, good agreement; and 0.81-1.0, excellent agreement. The analysis were performed using commercial statistical softwares (SPSS version 23, IBM Corporation, Armonk, NY or MedCalc ver. 16.4, MedCalc software, Marikerkem Belgium). For all statistical analysis, P value less than 0.05 were considered statistically significant.

Table 2 summarized important CT findings according to the resection margin status in perihilar cancer. By comparing histopathologic results, the AUC value of mid CBD involvement on pre-operative CT was 0.71 (cutoff score ≥ 4; sensitivity, 87.2%; specificity of 55.5%) and intrapancreatic CBD involvement was 0.72 (cutoff score ≥ 4; sensitivity, 60.0%; specificity of 91.6%). On univariable analysis, mid-CBD (score ≥ 4) or intrapancreatic CBD (score ≥ 4) involvement on pre-operative CT were significant factors for margin positive resection (p = 0.045 and 0.045, respectively). None of the vessel involvement, peritumoral fat infiltration, or lymphadenopathy on pre-operative CT were presented significant differences in resectability. On multivariable analysis, mid CBD involvement (score ≥ 4) on pre-operative CT (odds ratio = 3.09, 95% CI 1.44-6.67, p = 0.004) remained significant factor for margin positive resection (Table 2, Fig. 2). The mid CBD involvement (score ≥ 4) on pre-operative CT predicted 92.9% (26/28) of distal margin positive resection (p = 0.002), but did not associated with proximal margin (p = 0.254) or radial margin positive resection (p = 0.435).

In assessment of tumor extent, good to excellent interobserver agreement were found: intrapancreatic CBD (ICC = 0.83), mid-CBD (ICC = 0.73), first order of IHD (κ = 0.63) and second order of IHD (κ = 0.67) involvement. Regarding vessel relation assessment (κ = 0.32, 0.52 and 0.74 for RPV, LPV and MPV, κ = 0.35 and 0.52 for RHA and LHA, respectively) and peritumoral fat infiltration (ICC = 0.47), fair to good agreement were found.

Of 139 patients, 112 patients (80.6 %) were deceased owing to the following cause: progression of PHC (n = 101), pneumonia (n = 5), progression of other malignancies (n = 3), and sepsis (n = 3). The mean follow up time was 37.8 ± 34 months (95% confidence interval [CI], 19.2 - 31.4 months). The median survival was 25.3 ± 3.12 months (interquartile range, 11.5 - 55.9 months). The mean survival was 24.8 ± 19.5 months, and mean follow-up period for survivor was 92.0 ± 27.8 months.

Table 3 summarized important clinical and CT findings for overall survival after curative intent surgery in perihilar cancer. In univariable analysis, elevated level of total bilirubin (p = 0.03) and CA 19-9 (p = 0.02), intrapancreatic CBD (p < 0.01) and second order branches of IHD involvement (p = 0.07) on pre-operative CT were significant risk factor for survival. On multivariable analysis, increased level of total bilirubin (> 1.2 mg/dL , hazard ratio [HR] = 1.53, p = 0.04), CA 19-9 (> 37 U/mL, HR = 1.75, p < 0.01), and intrapancreatic CBD involvement (score ≥ 4, HR = 1.81, p < 0.01) on pre-operative CT were the significant predictor for poor overall survival (Figs. 3 and 4).