Study population
Institutional review board approval was obtained for this retrospective study, and the requirement for the informed consent was waived. From January 2012 to December 2016, a total of 590 consecutive patients who were pathologically diagnosed with HCC and underwent CECT examination within 2 weeks before curative resection were recruited. The inclusion criteria were as follows: (1) age > 18 years; (2) BCLC stage 0-B. The exclusion criteria were as follows: (1) previous treatment history (ablation, TKI or transarterial chemoembolization) (n = 39); (2) incomplete clinical data (n = 54); and (3) loss to follow-up or death without recurrence within 5 years after LR (n = 168). Ultimately, 329 patients were included in this study (Fig. 1). The clinical factors included age, sex, hepatitis B virus (HBV), Child–Pugh liver function, BCLC classification, largest tumor diameter, tumor number, microvascular invasion (MVI) and laboratory values [albumin (ALB), total bilirubin (TBIL), alanine aminotransferase (ALT), prothrombin time (PT), a-fetoprotein (AFP)]. The patients were randomly divided into training cohort (n = 230) and validation cohort (n = 99) at a ratio of 7:3. Figure 2 illustrated the workflow of our study.
Follow-up
Patients were followed up after LR, and HCC recurrence was screened by means of AFP and imaging examinations every 3 months. Recurrence-free survival (RFS) was defined as the time from the date of surgery to the date of first recurrence, metastasis, or last follow-up.
Image acquisition and preprocessing
CECT was performed using the following equipment: Toshiba Aquilion One, Phillips Brilliance iCT 256 (Philips Medical Systems) or Toshiba Aquilion 64 (Toshiba Medical Systems) scanners. The scanning parameters were as follows: tube voltage, 120 kVp; tube current, 200–250 mA; detector collimation, 64, 256 or 320 × 0.625 mm; field of view, 350 × 350 mm; matrix, 512 × 512; reconstructed slice thickness, 2 mm. After a routine unenhanced scan, we injected contrast material intravenously at a flow rate of 3.0 mL/s to obtain CECT images. Hepatic arterial phase CT images were obtained 30 s after injection. Through preprocessing, images were resampled to a volume of 1.0 mm*1.0 mm *1.0 mm to standardize the voxel spacing.
Tumor segmentation and feature extraction
Three-dimensional (3D) contours of the tumor regions of interest (ROIs) were delineated manually as gross-tumor regions (GTRs). All of the CT images were independently evaluated by two senior abdomen radiologists (9 and 15 years of experience) who were blinded to the clinical and pathological information. The radiologists delineated the boundaries of GTR on a transversal plane using ITK-SNAP (opensource software; version 3.4.0; www.itk-snap.org) software. The peritumor region (PTR) was defined as the parenchyma that fell within a 2-cm distance to the tumor boundary [18]. We outlined the tumor contours twice (each time for 1 cm) in three dimensions automatically using in-house software (Artificial Intelligence Kit, A.K., GE Healthcare) and modified the boundaries manually. Finally, two peritumor regions (PTR1 and PTR2) were acquired for each GTR. PTR1 (0–1 cm) represented a micrometastatic region; PTR2 (1–2 cm) represented the context of underlying cirrhosis. In patients with multiple (two or more) lesions, the largest lesion was segmented and used for subsequent analysis.
A total of 1188 radiomics features were generated automatically using A.K. from GTR and two PTRs (396 for each region). All radiomics features for each region could be divided into three groups: (a) morphological features (n = 9): describe the morphology of the ROI; (b) first-order features (intensity features, n = 42) : related to the distribution of the intensities of voxels within the ROI, ignoring the spatial interactions between them, and can be calculated from histogram analysis; and (c) texture features (n = 345) [19]: are able to quantify how pixels are positioned in relation to each other, e.g. grey level co-occurring matrix (GLCM) are co-occurring pixels in each defined direction and are counted and recorded into a matrix [20]. The details are provided in Fig. S1. Inter-/intraclass correlation coefficients (ICCs) were used to evaluate radiomics feature reliability [21]. To assess interobserver reproducibility, the ROIs of 30 randomly chosen images were performed by the two abdomen radiologists independently. To evaluate intra-observer reproducibility, the radiologists repeated the same procedure at one-month intervals. Features with ICC > 0.75 was considered to indicate good agreement and entered in next analysis.
Feature selection and radiomics model building
Least absolute shrinkage and selection operator Cox regression analysis (LASSO-Cox) was utilized to select the most significant radiomic features. LASSO has the key advantage of simultaneously selecting important variables and estimating their effects on the outcome from large number of candidate predictors. This selection method was proposed for linear or generalized linear models initially, where the outcome was fully observed. Then it has been extended to the Cox model (LASSO-Cox) for censored time-to-event responses. Based on the optimal lambda value that was selected through a 10-fold cross-validations, a panel of prognostic radiomic features was determined and built a GTR-PTR1-PTR2 model in the training cohort. The Rad-score in radiomics model was calculated via a linear combination of the selected features that had been weighted by their respective coefficients, represented quantitative ROI characteristics of each patient. Then, the GTR model and GTR-PTR1 model were established in the same manner, and their discriminative power was compared with the GTR-PTR1-PTR2 model.
Validation of radiomics model
-
(1)
High and low risk groups were classified according to the optimal Rad-score cut-point which was obtained by using X-tile software (version 3.6.1; Yale University School of Medicine, New Haven, Conn). Recurrence probabilities were estimated using the Kaplan-Meier method and compared by the log-rank test between the two subgroups.
-
(2)
The receiver operating characteristic (ROC) curves and areas under the curves (AUCs) were employed to investigate the performance at different RFS time points of 1-, 2- and 5-year.
Combined model construction and validation
Age, sex, hepatitis B virus (HBV), Child–Pugh liver function, BCLC classification, largest tumor diameter, tumor number, microvascular invasion (MVI) and laboratory values were analyzed by univariate Cox regression. Significant factors (p < 0.05) were chosen as candidates.
The combined model was constructed from the Rad-score and the candidate clinical factors by multivariate Cox regression; the significant factors with p < 0.05 were included in the model. A nomogram for 1-, 2- and 5-year RFS rate predictions was plotted according to the combined model. Then, the C-index and calibration curve were generated from the combined model.
Statistical analysis
Descriptive statistics are presented as the median (interquartile range [IQR]) for continuous variables and the frequency (%) for categorical variables. To determine significant differences between the training and validation cohorts, continuity correction and Pearson chi-square tests were used. All statistical tests were two tailed, and a p value < 0.05 indicated a significant difference. All statistical analyses were performed using R statistical software version 3.6.2.