Patients
Zhongshan Hospital, Fudan University ethics committee approved this retrospective- prospective study and the requirement of written informed consent was waived. From January 2016 to June 2019, 497 adult patients (age > 18 y.o.) were retrospectively collected and 249 patients (166 males and 83 females; 62.06 ± 10.37 years) with postoperative pathologically confirmed ICC lesions and preoperative Ga-DTPA-enhanced MRI examination were finally enrolled in accordance with the following inclusion criteria: (a) single primary liver lesion; (b) without prior history of anti-tumor treatment (including hepatectomy, transcatheter arterial chemoembolization, radiofrequency ablation, chemotherapy, radiotherapy, and immunosuppressive therapy, etc.); (c) complete histopathologic description of ICC lesion; (d) mass-forming ICC; (e) the interval between MRI examination and hepatectomy within 30 days; (f) with the longest diameter ≥ 1.0 cm, and without macrovascular invasion which can often be diagnosed with preoperative CT or MRI imaging; (g) high-quality MR images. All enrolled patients were randomly divided into training cohort (n = 174, 48 with MVI and 126 without MVI) and validation cohort (n = 75, 20 with MVI and 55 without MVI) in a ratio of 7:3. Critically, 69 ICC patients were prospectively collected and 47 ICC patients were finally enrolled as a time-independent test cohort (n = 47, 16 with MVI and 31 without MVI) from July 2019 to January 2020 in Zhongshan Hospital (Table S1). The flowchart of patient enrollment is presented in Fig. 1.
Clinical And Pathologic Data Evaluation
Clinical and laboratory variables include age, gender, hepatitis B virus (HBV) infection status, serum α-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 19 − 9 (CA199) within 7 days before curative resection. ICC specimens were collected using a 7-point baseline sample collection protocol(9). Pathologic data including MVI status, Edmondson-Steiner grade, and tumor number were cross confirmed by two experienced abdominal pathologists who have 5 and 13 years of experience in liver pathology, and was unaware of the results of MRI and clinical information. Consensus reached through discussion if there was any disputation. In current study, all ICC patients in three cohorts were divided into two subgroups according to the MVI status. Then, the baseline of clinical and laboratory variables in three cohorts were compared.
Gd-dtpa Mr Imaging Acquisition
Taking 3.0T uMR 770 scanner (United Imaging Healthcare, Shanghai, China) as an example. Non-contrast liver protocols consisted of axial T2-weighted imaging with fat suppression (T2WI-FS), diffusion-weighted imaging (DWI, with b values of 0, 50, 500 s/mm2), and axial pre-contrast quick three-dimensional T1-weighted imaging (quick3d T1WI) with fat suppression. Dynamic contrast-enhanced MR imaging was performed with T1WI-FS. Dynamic contrast-enhanced MR imaging performed by using gadopentetate dimeglumine (Primovist, Bayer HealthCare, Germany) for its functions to evaluate hepatic malignancies as both an extracellular agent and a hepatocyte-selected agent. After intravenous injection of (0.2 mmol/kg) of gadopentetate dimeglumine, followed immediately by a 20 ml saline flush at 2 ml/s, arterial phase (AP) images were acquired at 20–30s, portal venous phase (PVP) and delayed phase (DP) images were acquired at 60–70s and 180s, respectively. All the sequences and their detailed parameters could be found in Table S2.
Mr Imaging Feature Evaluation
All MR images were retrospectively or prospectively investigated on a picture archiving and communication system (PACS; Pathspeed, GE Medical Systems Integrated Imaging Solutions, Chicago, IL, USA) by 2 experienced abdominal radiologists (X.L.Q. and C.W.Z., who have 7 and 15 years of experience in abdominal imaging, respectively). They were blinded to all clinical and pathologic data. If any discrepancies arose between the two observers, a consensus was reached after discussion. Evaluation of MR imaging features as follows: (1) tumor size, defined as maximum tumor diameter on transverse MR T1WI image; (2) tumor morphology, including (hemi-)spherical/oval, lobulated and irregular; (3) signal intensity (SI) on T1WI, T2WI-FS, including hypointense, isointense, and hyperintense; (4) target sign on T2WI-FS and DWI, defined as a peripheral ring-like hyperintense with central isointensity/hypointensity(18); (5) arterial rim enhancement, defined as the hyperenhancement limited to peripheral of the lesion on AP, including complete and incomplete rim; (6) dynamic enhancement pattern: (A) gradual and filling; (B) arterial and persistent; (C)wash-in and wash-out; (7) the Liver Imaging Reporting and Data System (LI-RADS ver. 2018) categorization; (8) intrahepatic duct dilatation, including within and outside the lesion; (9) hepatic capsular retraction; (10) visible vessel penetration, defined as the presence of penetrating vessels in the lesion(19); (11) peritumoral enhancement, defined as hyperenhancement around the lesion compared with the background liver parenchyma around the lesion on any phase.
Radiomics Analysis
Workflow
The workflow of the radiomics analysis in our study includes tumor segmentation, feature extraction, feature selection, model construction and model evaluation (Fig. 2).
Tumor Segmentation
The volumes of interest (VOI) were manually delineated in ITK-SNAP (ver. 3.6.0) (http://www.itksnap.org/pmwiki/pmwiki.php) on six sequences: DWI with b value of 500 s/mm2, T2WI-FS, pre-T1WI, AP, PVP, and DP by two abdominal radiologists with 3 and 7 years of experience (G.Y.M. and X.L.Q.) and validated by a senior abdominal radiologist with 17 years of experience (X.L.).
Feature Extraction
Radiomics features were extracted by using PyRadiomics (Version: 3.0.1; https://pyradiomics.readthedocs.io/en/latest/). In total, 1688 radiomics features per sequence were extracted, which are grouped into the following classes: First Order Statistics, Shape-based (3D), Shape-based (2D), Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Matrix (NGTDM) and Gray-Level Dependence Matrix (GLDM) features.
Feature Selection
In this study, firstly, a total of 6×1,688 = 10,128 extracted features from 6 MRI sequences were applied with a z-score normalization to eliminate index dimension difference. Secondly, we conducted a repeat segmentation procedure on 30 randomly selected ICC lesions, and 9,175 features with intraclass correlation coefficients ≥ 0.9 in test-retest were considered as stable features. Thirdly, after further variance filtering correlation analysis, we obtained 5,962 features. Then, 940 features were obtained by Student's t test (p ≤ 0.05). Finally, we obtained 25 stable MRI-based radiomics features associated with MVI status by LASSO regression (Table S3, Fig.S1). And the detailed information of 25 radiomics features is listed in Table S4.
Model Construction
The selection of clinical and imaging features was performed by applying the univariable logistic regression, and multivariable logistic regression analysis was used to elucidate the independent features of MVI status. The clinical or/and imaging model is produced by a linear combination of independent clinical or/and imaging features, with their respective coefficients weighted.
Radiomics model was firstly constructed by Decision Tree (DT), K-Nearest Neighbor (KNN) algorithm, Logistic Regression (LR), Random Forest (RF) and Support Vector Machine (SVM) classifier respectively. Secondly, after comparison of the predictive performance among the five radiomics models, the optimal model with the greatest diagnostic potential was picked as the final radiomics model. Finally, an optimal cut-off value of radiomics score (Rad-score) was derived by plotting the ROC curve and calculating the maximum Youden index. By combining the above two models, we further obtain the fusion model.
Model Evaluation
The receiver operating characteristic curves (ROC) of each model in three cohorts were plotted, and AUC were used to quantify the predictive efficacy of models, and corresponding accuracy, sensitivity and specificity were also calculated. Calibration curves as visual tools were plotted to evaluate the agreement between the predicted MVI status and the actual MVI status by the Hosmer-Lemeshow test. Decision curves of each model in three cohorts were plotted to assess the clinical net benefit at different risk thresholds by decision curve analysis.
Follow-up
296 patients were postoperatively followed up by the measurement of serologic tumor markers, abdominal ultrasonography, contrast-enhanced CT or MRI at intervals of 3 to 6 months after hepatectomy, but 7 patients were finally lost to follow-up. Recurrence was determined in case of unequivocal new tumor detected by CT/MRI or pathological confirmation. The last follow-up date for this cohort was October 2022. OS was calculated from the date of the surgery to the date of the death or the last follow-up.
Overall Survival Analysis
Overall survival analysis was constructed in histological MVI group and predicted MVI group. Histological MVI-positive subgroup and predicted MVI-positive subgroup were defined as high-risk subgroups. And histological MVI-negative subgroup and predicted MVI-negative subgroup were defined as low-risk subgroups. Univariate and multivariate Cox regression were applied to identify the preoperatively independent risk factors of overall survival in histological and predicted MVI groups. Notably, correlation analysis was conducted between the univariate and multivariate Cox regression, correlation coefficients over than 0.7 indicating strong correlation, variable with strong correlation would not be included in the subsequent multivariate Cox regression(19, 20). Survival probabilities were evaluated and plotted by the Kaplan–Meier (KM) survival analysis, and compared by the log-rank test. Finally, nomograms of histological and predicted MVI groups were also plotted to anticipate the probability of 1-year, 3-year, and 5-year OS.
Statistical analysis
For continuous variables, Student t test was used when variables with normal distribution, and Mann-Whitney U test was used when variables with non-normal distribution. For qualitative variables, Chi-square test or Fisher exact test was used. Considering the increased statistical utility and reduced potential model overfitting, features that were too rare or too common were excluded before performing logistic regression and Cox regression analysis(21). Univariable and multivariable logistic regression were used to elucidate the independent clinical and imaging features of MVI status and the independent clinical and imaging features of OS status were identified by univariable and multivariable Cox regression. The predictive efficacy between multiple models were compared by the Delong test. All statistical analysis and curves were conducted and plotted by using SPSS (version 21.0; IBM) and R software (Version 3.4.1). All tests were two-tailed and p-value of less than 0.05 was considered statistically significant.