Prediction for early recurrence of intrahepatic mass-forming cholangiocarcinoma: quantitative magnetic resonance imaging combined with prognostic immunohistochemical markers

This retrospective study was approved by our Institutional Review Board, and the need for informed consent was waived. We collected electronic medical records from May 2011 to August 2016 at our institution. A total of 219 patients who underwent a curative-intent resection and lymph node dissection, with histopathologically confirmed IMCC, were recruited. Among these patients, the study population was selected using the following inclusion criteria: (a) patients who underwent preoperative liver CE-MRI within 4 weeks of their surgery, (b) patients without a history of previous adjuvant treatment before the surgery, (c) patients with histopathologically proven IMCC and negative resection margin (R0), combined hepatocellular-cholangiocarcinoma were excluded, (d) patients without a history of other tumors, and (e) patients who completed at least 2 years of follow-up. Consequently, 47 patients were enrolled in our study and were divided into an ER group (n = 31) and a non-ER group (n = 16), with ER being defined as the development of intrahepatic or extrahepatic recurrence within 2 postoperative years (Fig. 1).

Clinical and pathological characteristics consisted of age, gender, hepatitis, carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA199), satellite lesions, maximum tumor diameter (MTD), tumor location, differential degree of tumor, stage and lymph node metastases. Lymph node metastasis was determined by the pathological results of lymph node dissection and preoperative CT/MR imaging. The threshold values chosen for CA199 and CEA levels were based on the normal ranges used at our institution (0-37 U/ml for CA199 and 0-5 ng/ml for CEA).

All pathology sections and macroscopic pictures of the resected specimen were retrospectively reviewed. Combining with immunohistochemistry, all 47 patients were confirmed IMCC by post-operative pathology. Immunohistochemical staining was detected on formalin-fixed, paraffin-embedded sections using standard immunohistochemical methods. Then five-micron-thick sections were created, and antibodies specific for EGFR, VEGFR, P53, and Ki67 (Beijing Zhongshan Golden Bridge Biotechnology Co. LTD, China) were used to perform the further immunohistochemical staining. All samples were analyzed by an anatomic pathologist with 10 years of experience, who was unaware of the patient’s outcome. Less than 10% of the positive staining was identified as negative expression, while more than 10% of the positive staining was identified as positive expression.

All patients underwent contrast enhanced CT or MRI every 3–6 months after surgery in the first 2 years. Images were analyzed to identify ER, which was determined as the presence of new intrahepatic lesions with typical imaging features of IMCC, atypical lesions with histopathological confirmation, or extrahepatic metastasis (lymph node metastases or distant metastasis) within 2 postoperative years.

All patients underwent 3.0 T MRI scans (Signa Excite HDxt, GE Healthcare, Milwaukee, USA) with an eight-element phased-array torso coil. After nonenhanced axial breath-hold T1 weighted imaging, fat suppression T2 weighted imaging (T2WI/FS) and DWI (b-values of 0 and 800 s/mm²), contrast enhanced T1-weighted three-dimensional (3D) spoiled gradient echo sequence (LAVA) was performed. Gadodiamide (Omniscan 0.5 mmol/ml; GE Healthcare, Ireland) was injected at a dose of 0.2 mL per kilogram and a rate of 2 mL per second as a bolus by an automatic pump injector and a subsequent 20 mL 0.9% sterile saline flush. Contrast enhanced imaging was performed in the arterial phase (AP) (30 s), portal venous phase (PVP) (60 s), and delayed phase (DP) (180 s). Images of AP, PVP, and DP were all used for radiomics feature extraction and analysis. Additional technical details are provided in Table 1.

MR images were reviewed independently by two radiologists who had 5 and 10 years experience in the abdominal MRI, respectively. Both radiologists were blinded to the clinical data of the patients when they evaluated the MR images. They reached a consensus by discussion when there were disagreements. The basic imaging traits potentially associated with ER included lesion shape, contour, biliary dilation, capsular retraction, DWI intensity and the enhancement pattern. The target appearance was peripheral hyperintensity compared to the center on high b-value DWI. The enhancement pattern was assessed using the following subdivisions: (a) gradual enhancement (the enhancement area gradually increased from the periphery to the center of the tumor), (b) persistent enhancement (enhancement remained through all phases), (c) wash in and wash out (hyperenhancement of the AP followed by washout), and (d) minimal or no enhancement.

MR images (AP, PVP, DP, T2WI/FS sequence) were loaded into ITK-SNAP software (version 2.2.0, www.​itksnap.​org) for 3D manual segmentation. A radiologist with 10 years of MRI experience (reader 1) performed the tumor segmentations in all 47 patients. After 2 weeks, images of all patients were segmented again by reader 1 and another radiologist (reader 2) with 5 years of experience of MRI diagnosis to assess intra−/inter-reader agreement in the feature analysis. All outcomes were based on the features extracted by the first segmentation from reader 1.

Artificial Intelligence Kit software (A.K. software; GE Healthcare, Life Sciences, Beijing, China) was used to extract 396 parameters from each sequence. Those parameters include first order histogram features (n = 42), grey-level co-occurrence matrix (GLCM) features (n = 144), grey-level run-length matrix features (n = 180), Haralick features (n = 10), morphological features (n = 9) and grey-level zone size matrix features (n = 11).

The proposed parameters were analyzed for consistency and correlation. First, the intraclass correlation coefficient was determined for each parameter for the inter-observer and intra-observer reproducibility test. Features with intraclass correlation coefficient values less than 0.8 were excluded. Second, stratified analyses were conducted using the Wilcoxon signed-rank test to discover the potential association between the remaining parameters and ER status, followed by univariate logistic regression. To keep discriminative parameters, we set a threshold of 0.1. A variance inflation factor was then used to eliminate parameters with high collinearity in a multiple mutual linear situation. Finally, multivariate logistic regression was applied to evaluate the performance of distinguishing ER status in each sequence. Different sequence combinations (AP + PVP, AP + PVP + DP, AP + PVP + DP + T2WI) were also tried to explore the best model using the same methods described above. The validation of each model was performed by using leave-one-out cross-validation. Thus, three predictive models were built: a best performance radiomics model, a clinicoradiologic-pathologic(CRP) model, and a combined model with both selected radiomics features and clinicoradiologic-pathologic features (Fig. 2).

The differences in patient characteristics between the two groups were assessed using t-test or the Mann-Whitney U test for continuous variables and the chi-square test or Fisher exact test for categorical variables. Kappa tests were used to determine inter-observer agreement for qualitative MRI features. Kappa values of 0.81–1.00 indicated excellent agreement, 0.61–0.80 signified substantial agreement, and 0.41–0.60 denoted moderate agreement. Receiver operating characteristics (ROC) curves were performed in each model. The area under the curve (AUC) of the ROC curves, accuracy, sensitivity, specificity, positive predictive values (PPVs) and negative predictive values (NPVs) were obtained and comparisons between the performance of final three models were performed using the Delong test. All statistical analyses were performed using R studio Server (Version 3.5.0; RStudio, Inc., Boston, MA, USA) and SPSS (version 20.0; IBM, Armonk, NY, USA). A two-sided P value less than 0.05 indicated a statistically significant difference.