Patients
From January 2012 to December 2020, a total of 1267 patients with HCC underwent hepatic resection in our hospital. Of these, 254 patients with huge HCC who underwent curative surgical resection were recruited. However, 68 patients were excluded, and 186 patients who met the following inclusion criteria, were enrolled into this study: (1) patients who did not receive any treatment before surgery; (2) Liver function was classified as Child-Pugh grade A or B; (3) Eastern Cooperative Oncology Group (ECOG) performance score 0-2; (4) Undergone enhanced CT scanning within 7 days before surgery; (5) histologically confirmed HCC. The detailed recruitment process is presented in Fig. 1. Then, patients were divided into training and test datasets at a ratio of 7:3. The training dataset was used to construct prediction model, and the test dataset was used to validate the performance of model. The Ethics Review Board of the Second Affiliated Hospital of Zhejiang University School of Medicine approved this study (No. 2021-0376).
Clinical Characteristics
Baseline demographic, clinical, and laboratory characteristics (including liver and kidney function tests, platelet count, blood coagulation index, serum alpha-fetoprotein level), clinical grading scores were collected. The operative variables (including surgical methods, intraoperative blood loss, intraoperative blood transfusion, and intraoperative vascular occlusion methods) that correlate to PHLF were also recorded.
Diagnosis and definitions
PHLF was diagnosed according to the International Study Group of Liver Surgery (ISGLS) criteria [17]. The INR was set at 1.5 and the bilirubin level of more than 20 micromol/L (1.2mg/dL). The severity of PHLF was divided into 3-classes according to the clinical management: grade A, no further clinical management necessary; grade B, requires an active therapeutic intervention without invasive approach; grade C, invasive approach. We defined grade B and C PHLF as severe PHLF and put as primary outcome since grade A PHLF does not require any additional management.
CT scan acquisition
Multi-detector CT systems (16-slice SOMATOM Perspective, SIEMENS; 16-slice SOMATOM Sensation, SIEMENS, Germany) were scanning devices in our department. Dynamic contrast-enhanced CT imaging were obtained following the administration of iodinated contrast material (Iohexol, GE Healthcare, USA) at 3.0 mL/sec. Scanning parameters were 120 KV, 160 mAs; rotation time 0.5 s; 350 mm×350 mm field of view; matrix of 388×388; slice thickness, 3 mm. Arterial phase and portal phase images were obtained at 40s and 72s after injection of contrast medium.
Image segmentation and radiomics features extraction
The region of interest (ROI) was drawn manually using the freely available application ITK-SNAP (version 3.6.0). ROI was delineated in the liver along the border of the whole liver parenchyma by avoiding major blood vessels, focal lesions, and artifacts on the portal phase images. Features were extracted from each segmented ROI, divided into textual and non-textural features using PyRadiomics [18], an open-source python package for medical imaging.
To get reproducible radiomics features, standardized computation of radiomics features is necessary [19]. In our study, the sitkBSpline interpolation was applied to resample the images with a pixel size of 1×1mm. Voxel intensities were discretized
using a bin-width of 25 HU. 788 radiomics features were extracted from the liver ROI, including 18 original first-order histogram features, 14 original shape features, 68 original textural features and 688 high-order wavelet features. The list of radiomics features is shown in the supplemental table 1.
Inter-observer and intra-observer agreement
To ensure reproducibility, CT images of 20 patients were randomly selected and re-segmented by reader 1 (X.F. with 7 years’ experience in liver imaging) at an interval of 2 week and reader 2 (Y.L.L. with 8 years’ experience in liver imaging) independently performed the segmentation. The intra-observer reproducibility and inter-observer reliability of features extraction were assessed by using intra- and inter-class correlation coefficients (ICCs). Features with ICC > 0.75 represent a good agreement and were retained.
Feature selection and radiomics signature construction
The extracted radiomics features were normalized by Z-scores method. Radiomics features with ICCs lower than 0.75 were excluded. Univariate analyses were conducted using univariate logistic regression analysis. Features were considered to be associated with severe PHLF when p values < 0.1. Based on the selected features, the least absolute shrinkage and selection operator (LASSO) algorithm was conducted to identify significant features with nonzero coefficients. The penalty parameter (𝜆) was optimized through the tenfold cross-validation method. A radiomics signature was constructed by summing the selected features multiplied by their coefficients. The area under the receiver operating characteristic curve (AUC area under the ROC curve) was calculated for assessing the predictive ability of the established radiomics signature.
Development of the clinical-radiomic nomogram
To develop a comprehensive clinical-radiomic nomogram, the clinical characteristics and radiomics signature were analyzed by univariate logistic regression. Significant factors (p<0.05) were used to build the multivariate logistic model. Finally, the clinical-radiomic nomogram model integrated the clinical predictors and radiomics signature was built in the training dataset.
Assessing the accuracy of Nomogram model and comparison with conventional methods
We determined the discriminatory ability of the nomogram model, by comparing the radiomics signature, albumin-bilirubin score (ALBI) score, the model for end-stage liver disease (MELD) score, Child-Pugh score with the areas under the receiver operating characteristic curve (AUC). DeLong's test was used to compare the nomogram model with conventional methods based on the AUC values in both datasets. To evaluate the consistency of the nomogram, we plotted a calibration curve with the Hosmer-Lemeshow goodness-of-fit test.
Clinical Use
To assist in surgical decision-making, a decision tree for safe huge HCC hepatectomy was built based on the associated risk factors.In addition, to evaluate the clinical usefulness of the nomogram model, radiomics signature, MELD, ALBI, and Child-Pugh scores, decision curve analysis (DCA) was conducted to evaluate the net benefits across a variety of threshold risks.
Statistical Analysis
The radiomics analysis workflow is shown in Fig. 2. Continuous variables and categorical variables were compared by Mann–Whitney U test and chi-square test, respectively. We considered two-tailed values of p < 0.05 as statistically significant for all analyses. All analyses were conducted with R software (version 3.6.1).