This retrospective study was conducted at a single center and was approved by the Ethics Committee of West China Hospital, Sichuan University (Approval No. 2022–651). The requirement for written informed consent was waived due to the retrospective nature of this study.

From January 2010 to May 2022, 334 eligible patients (283 males) with a mean age of 53.0 ± 11.6 years were included in this study (Fig. 1), who met the following inclusion criteria: (a) no less than 18 years of age, (b) underwent curative hepatic resection, (c) had pathologically confirmed HCC, (d) underwent contrast-enhanced MRI within 30 days before surgery, and (e) underwent complete immunohistochemical staining for CK7 and CK19 which was included on the postoperative pathology report. The exclusion criteria were as follows: (a) received non-curative liver resection; (b) had multiple HCC (HCCs equal to or more than two); (c) received any antitumor treatment for HCC prior to surgery; (d) had insufficient MR imaging quality (e.g., severe artifact); and (e) presence of malignant tumors other than HCC (for example, combined hepatocellular-cholangiocarcinoma [cHCC-CCA] and sarcomatoid carcinoma, cHCC-CCA contains areas of both typical HCC and typical iCCA, the former having any/all of the possible cytological and architectural features of HCCs and the latter distinctly being an adenocarcinoma with malignant glands, usually lying within a dense stromal background). Details of patient inclusion and exclusion criteria are presented in Fig. 1.

Flowchart of the retrospective study cohort. A total of 334 patients diagnosed with solitary hepatocellular carcinoma were included in this research

Baseline clinical information, including patient demographics, causes of liver disease, and key laboratory test results (alpha-fetoprotein [AFP] and carbohydrate antigen 199 [CA199]) within 14 days of surgery, were obtained from electronic medical records.

Four 3.0-T MR scanners (Discovery 750, SIGNA™ Architect and SIGNA™ Premier, GE Healthcare; and MAGNETOM Skyra, Siemens Healthineers) and one 1.5-T MRI scanner (uMR588, United Imaging Healthcare) were used to acquire MR images. The sequences employed T2-weighted imaging, diffusion-weighted imaging, T1-weighted in-phase, and opposed-phase imaging, as well as T1-weighted dynamic contrast-enhanced imaging using gadopentetic acid dimeglumine or gadoxetic acid disodium (Primovist®, Bayer Pharma AG). Supplementary A1 and Table S1 provide detailed information regarding the MRI techniques.

Two radiologists with 8 and 6 years of experience in liver MR, respectively, independently assessed all MR scans. Although the reviewers were aware that all patients had HCC, they remained blinded to other clinical, histopathological, and follow-up information. In cases where there was a discrepancy in image interpretation, a third radiologist with over 20 years of experience in liver MR assessed the image to provide a resolution.

A total of 24 pre-operative MRI features were assessed. These features encompass those related to the underlying liver disease (e.g., radiologically evident cirrhosis) and other prognostic features (e.g., intratumoral artery, tumor growth subtype, non-smooth tumor margin, and peritumoral hepatobiliary phase hypointensity) of HCC. Descriptions of the assessed features are summarized in Supplementary Table (S2).

Data on tumor location, number, size, Edmondson-Steiner differentiation grade, immunohistochemical expression of CK7 and CK19, hepatocyte paraffin antigen 1 (HepPar-1), glypican-3 (GPC-3), glutamine synthetase (GS), and microvascular invasion were collected from pathology reports. The expression of CK7, CK19, HepPar-1, GPC-3, and GS was classified as negative or positive. All histopathological examinations were conducted by two pathologists (with over 5 and 10 years of experience in liver pathology) who were blinded to the clinical and imaging information. HCC with the cholangiocyte phenotype was pathologically diagnosed if all of the following criteria were fulfilled: (I) microscopic morphological features of HCC; (II) positive expression of HepPar-1, GPC-3, or GS in tumor cells; and (III) positive expression of CK7 and/or CK19 in tumor cells (≥ 15%) [5, 6, 14].

After surgery, patients were followed up at 1 month, 2 months, and then every 3 months for the first 2 years. Subsequently, follow-ups were performed every 6 months. During each follow-up, serum AFP levels were measured, and contrast-enhanced ultrasound, CT, or MR imaging was performed. Additionally, tumor recurrence was confirmed by imaging or pathological examination during the follow-up period. The administration of adjuvant therapy (e.g., systemic therapy and transcatheter arterial chemoembolization) prior to recurrence had been documented. Recurrence-free survival (RFS) was defined as the duration from the date of surgery to the occurrence of tumor recurrence or the last follow-up date (May 1, 2022), whichever occurred first. Patients who died from causes unrelated to tumor recurrence were censored without an event when calculating RFS. Overall survival (OS) was defined as the duration from the date of surgery to the date of death from any cause or the last follow-up date, whichever occurred first.

A predictive model for cholangiocyte phenotype HCC was developed and validated. According to the chronological order of MRI examinations, those patients included in this study were divided into a model training dataset (232 patients) and an external validation dataset (102 patients) in a 7:3 ratio.

In the training dataset, univariate logistic regression analyses were conducted to identify clinicoradiological features associated with the cholangiocyte phenotype of HCC. Continuous variables were converted into categorical or dichotomized variables based on normality ranges or clinical relevance to enhance their clinical applicability. Multicollinearity was assessed using the variance inflation factor. All independent predictors with p-values < 0.1 in the univariate analyses were included in the multivariate logistic regression model, which utilized the backward stepwise method and fivefold cross-validation to create an “internal validation” dataset. Patient age, sex, and hepatitis B virus infection status (infected vs. non-infected) were controlled for in the model. Akaike Information Criterion was used to obtain the most parsimonious feature combination. Therefore, these features were selected because their combination allowed the lowest Akaike Information Criterion among all feature combinations. However, this approach did not correspond to all p < 0.05 [15, 16]. A scoring system was developed using the predictors identified in the multivariate regression analysis to estimate the probability of the cholangiocyte phenotype in HCC. The optimal threshold of the scoring system was determined using the receiver operating characteristic (ROC) curve analysis and Youden's index.

A ROC curve was used to compute the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy of the discriminative performance of the model. Calibration curves were plotted to assess the calibration of the model using the Hosmer–Lemeshow test. Furthermore, decision curve analysis was conducted to evaluate the clinical utility of the model by quantifying the net benefits at various threshold probabilities.

The Shapiro–Wilk test was used to assess normal distributions. Differences in continuous variables were analyzed using either the independent samples t-test or the Mann–Whitney U test. Categorical variables were evaluated using the chi-square test or Fisher’s exact test.

Cohen’s κ values or weighted κ values were used to evaluate the inter-rater agreement between the two reviewers in the MRI analysis.

The Kaplan–Meier technique and log-rank test were used to evaluate survival outcomes. Patients were excluded from the survival analysis if they had received systemic therapy (n = 31) or transcatheter arterial chemoembolization (n = 20) before recurrence, if tumor recurrence or death occurred within 2 months after surgery (n = 11), and if follow-up data were incomplete (n = 121).

All statistical analyses were conducted using the R project for statistical computation (version 4.0.5). Statistical significance was set at p < 0.05.