We retrospectively enrolled 537 HCC patients from two institutions (Affiliated Hospital of Guilin Medical University; Peking University People’s Hospital) who underwent radical resection. According to the exclusion criteria in Figure 1, 277 patients enrolled from October 2009 to May 2017 at the Affiliated Hospital of Guilin Medical University were set as the training cohort, while 125 patients enrolled from June 2010 to December 2017 at Peking University People's Hospital were set as the validation cohort. Along with the CECT images, information regarding demographic characteristics, clinicopathological data, and laboratory data were also retrospectively collected. Patient demographic characteristics included sex and age, while clinicopathological data were collected from electronic case records, containing tumour size and vascular invasion. Furthermore, laboratory data were collected before surgery, including routine blood test, liver function, and alpha-fetoprotein (AFP) level.

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Flowchart of the study cohorts. HCC: Hepatocellular carcinoma; CT: Computed tomography; TACE: Transarterial chemoembolization; ROI: Region of interest.

All enrolled patients were diagnosed with HCC by postoperative pathological examination; solitary HCC was confirmed by preoperative CECT and intraoperative palpation or US. In addition to CECT, at least US and MRI supported an HCC diagnosis. Radical resection was defined as complete removal of the tumour with no residual tumour or new lesion observed in two observations at an interval of no less than 4 wk. Two independent pathologists histopathologically examined resected tumour specimens. The main exclusion criteria were as follows: (1) CECT image unavailable; (2) CECT scan over 1 mo before operation; (3) previous treatment including hepatectomy, ablation, or transarterial chemoembolization; and (4) incomplete clinical or follow-up data.

Postoperative surveillance at each institution was conducted according to the protocol. Patients were followed every 2 mo within the first 2 years and every 3-6 mo after 2 years postoperatively (abdominal ultrasonography, serum AFP, and other serological tests were performed). CECT or MRI examination was performed if recurrence was suspected. OS was determined as the interval between the operation date and the date of death or last follow-up. Recurrence-free survival (RFS) was determined from the operation date of radical surgery to the date of the first recurrence at any site, death, or the last follow-up. Recurrence within 2 years after operation was considered as early recurrence.

The CECT images were obtained from each institution according to routine procedures, and the usage of contrast medium is described in the Supplementary material.

Regions of interest (ROIs) were defined as the tumour areas in both arterial and portal venous phases, which were delineated via 3D Slicer (v4.11, https://www.slicer.org) by two experienced radiologists (Reader 1: H.Y. with 10 years of working experience; reader 2: Z.Z. with 20 years of working experience). If they could not reach a consensus, the third radiologist intervened in the evaluation.

To minimize the variability and normalize the CT images, all the images were resampled to a voxel size of 1 mm × 1 mm × 1 mm. Then, the radiomics features were extracted from the ROI segments based on the Pyradiomics package (version 3.0.1). The categories of radiomics features were as follows: First-order statistics, shape-based features (2D and 3D), grey level cooccurrence matrix (GLCM), grey level run length matrix (GLRLM), grey level size zone matrix (GLZM), neighbouring grey tone difference matrix (NGTDM), and grey level dependence matrix (GLDM). LASSO Cox regression was performed to filter and obtain the early recurrence-related radiomics features according to the minimum criteria with 10-fold cross-validation. The radiomics score was determined according to the coefficients.

A radiomics and clinical combined model was developed via univariate and multivariate analyses, in which the optimal cut-off value of the radiomics score was determined via X-tile (Version 3.6.1). The candidate clinical indicators were hepatitis B surface antigen (HBsAg) (positive vs negative), cirrhosis (present vs absent), microvascular invasion (MVI) (present vs absent), serum AFP (> 200 vs ≤ 200 ng/mL), tumour size (cm), platelets (× 10⁹/L), gamma-glutamyl transpeptidase (GGT) (U/L), and ALBI. The combined model was presented as a nomogram and validated in an independent cohort.

Continuous variables that follow a normal distribution are shown as the mean ± SD and were compared by using the Student’s t test. Otherwise, the Mann-Whitney U test was performed. Categorical variables were compared by chi-squared tests. RFS and OS analyses were conducted by using the Kaplan-Meier method and compared by log-rank test among different groups. Time-dependent receiver operating characteristic (ROC) curves were plotted based on the timeROC package. The concordance index (C-index) was determined to evaluate the predictive ability of the combined model. The calibration curves were plotted with the rms package, while decision curve analysis (DCA) was based on the rmda package. SPSS18.0 (SPSS Inc., Chicago, IL, United States) and R (version 4.0.3, https://www.rproject.org/) were used for statistical analyses. P < 0.05 was considered statistically significant.

A total of 402 patients who underwent preoperative CECT were enrolled according to the exclusion criteria from two institutions, and the baseline characteristics are illustrated in Table 1. There were no significant differences in the variables evaluated between the two cohorts except aspartate aminotransferase (AST); most patients were male (both 85.60%), with liver cirrhosis (92.06% and 93.60%) and HBV-related HCC (83.75% and 82.40%, respectively). In institution 1, the median follow-up time was 54.2 mo, the median OS time was 40.0 mo, and the early recurrence rate was 36.82% (102 recurred among 277). In institution 2, the median follow-up time was 50.7 mo, the OS time was 37.0 mo, and the early recurrence rate was 32.80% (41 recurred among 125).

A total of 1915 CECT-based radiomics features were extracted from images acquired at both the arterial and portal venous phases among 277 patients in institution 1 as the training cohort. To filter early recurrence-related features, LASSO Cox regression was performed, and 31 features were obtained (Supplementary Figure 1). Figure 2 displays the distribution and correlations among these features. Then, a radiomics score was determined according to the coefficients in Supplementary Table 1 for each patient, and the optimal cut-off value was identified as 15.93, based on which the low and high radiomics risk groups were divided accordingly. As shown in Figure 3A, the radiomics scores were significantly different between the early and nonearly recurrence groups, while the area under curve (AUC) value was 0.73 for both institutions (Figure 3B), thus demonstrating the predictive value of the radiomics score for early recurrence HCC.

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Correlation and distribution of early recurrence-related radiomics features. A: Correlation of the radiomics features. The red mark represents a positive correlation, while the blue mark represents a negative correlation; B: Heatmap and clustering analysis of the radiomics features. ROC: Receiver operating characteristic; RFS: Recurrence-free survival; AUC: Area under curve.

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Capacity of the radiomics score to distinguish early recurrence. A: Correlation of the rad-score between the nonearly recurrence and early recurrence groups; B: Receiver operating characteristic curves of the radiomics score to predict recurrence-free survival in the training and validation cohorts.

To further improve the predictive performance of the model, univariate and multivariate Cox regression analyses were conducted among the radiomics model and clinicopathologic indicators. The results from univariate analysis (Table 2) illustrated that the radiomics score, MVI, serum AFP, tumour size, and GGT were statistically significant and were included in the multivariate analysis. Multivariate analysis showed that radiomics score [hazard ratio (HR), 3.86; 95%CI: 2.41-6.16; P < 0.001] and serum AFP (HR, 1.52; 95%CI: 1.00-2.30; P = 0.048) were identified as independent predictive indicators for early HCC recurrence.

Next, a predictive and visible nomogram was developed based on the radiomics and clinical combined model (Figure 4), and the AUC values of the combined model were 0.79 (95%CI: 0.73-0.84) and 0.77 (95%CI: 0.70-0.83) for 1-year and 2-year recurrence prediction in the training cohort (Figure 5A), respectively. The AUC values of the validation cohort were 0.65 (95%CI: 0.58-0.73) and 0.74 (95%CI: 0.63-0.77) for 1-year and 2-year recurrence, respectively (Figure 5B). The C-index was 0.712 and 0.674 in the training and validation cohorts, respectively. Moreover, the calibration curves illustrated good agreement between predictive and observed outcomes (Figure 5C and D), while the DCA showed that the combined model had a relative high net benefit (Figure 5E and F).

Open in New Tab   Full Size Figure   Download Figure

Figure 4 A radiomics and clinical combined model is built based on the radiomics score and preoperative alpha-fetoprotein. According to the total score of each indicator, the recurrence rate corresponding to the total score is the nomogram-predicted rate. AFP: Alpha-fetoprotein; RFS: Recurrence-free survival.

Open in New Tab   Full Size Figure   Download Figure

Figure 5 Evaluation of the radiomics model to predict early recurrence. A: Receiver operating characteristic (ROC) curves to predict 1- and 2-year recurrence-free survival (RFS) in the training cohort; B: ROC curves to predict 1- and 2-year RFS in the validation cohort; C: Calibration curves of the 1- and 2-year RFS in the training cohort; D: Calibration curves of the 1- and 2-year RFS in the validation cohort; E: Decision curve analysis of the combined model in the training cohort; F: Decision curve analysis of the combined model in the validation cohort. ROC: Receiver operating characteristic; RFS: Recurrence-free survival; AUC: Area under curve.

To verify the discrimination ability of the combined model, survival curves for RFS and OS were plotted based on the median risk points in the training and validation cohorts. For early recurrence, the high-risk group was more likely to present recurrence than the low-risk group in both the training and validation cohorts (log-rank, P < 0.0001 for both) (Figure 6A and C). Furthermore, OS was significantly different (log-rank, P < 0.0001 for the training cohort and P = 0.0446 for the validation cohort) (Figure 6B and D).

Open in New Tab   Full Size Figure   Download Figure

Figure 6 Cumulative rates of early recurrence and survival rates of hepatocellular carcinoma. A: Risk accumulation curve of recurrence-free survival (RFS) in the training cohort; B: Survival curve of overall survival (OS) in the training cohort; C: Risk accumulation curve of RFS in the validation cohort; D: Survival curve of OS in the validation cohort; E and F: Risk accumulation curve of RFS in the alpha-fetoprotein (AFP)-negative group in the training and validation cohorts; G and H: Risk accumulation curve of RFS in the AFP-positive group in the training and validation cohorts. AFP: Alpha fetoprotein.

Next, we performed subgroup analysis in AFP-negative (< 20 ng/mL) and AFP-positive (≥ 20 ng/mL) groups. Regardless of the AFP level (Figure 6E-H), the model calculated points were closely related to early recurrence and OS (log-rank, P < 0.0001 for all). The results above illustrated the discrimination ability of the combined model.

Two representative case reports were displayed to demonstrate the distinguishing ability for similar cases (Figure 7). Both of them were male with BCLC A stage HCC and similar ages, and the ROI, also known as the tumour area, was similar in tumour size and position. MVI was present in AFP-negative case 1, while MVI was absent in AFP-positive case 2. However, their outcomes were different. Case 1 was both low risk at the radiomics score and model prediction point and without recurrence until the last follow-up. Case 2 was high risk with worse outcomes, and recurrence was present at 14 mo after the operation. Thus, we concluded that the radiomics combined model could support to distinguish similar cases using CECT in clinical practice.

Open in New Tab   Full Size Figure   Download Figure

Figure 7 Representative case reports. AFP: Alpha-fetoprotein; BCLC: Barcelona Clinic Liver Cancer; CECT: Contrast-enhanced computed tomography; MVI: Microvascular invasion; ROI: Regions of interest.

Hepatocellular carcinoma (HCC) seriously endangers human life and health, but there is still a lack of satisfactory treatment options. Even if it is diagnosed at early stage, the recurrence rate is still very high. The clinical monitoring strategy for HCC recurrence is limited, so there is a need to find a new and effective recurrence prediction model for HCC. And we developed a radiomics model based on preoperative contrast-enhanced computed tomography (CECT) to evaluate early recurrence in patients with a single tumour.

Due to the high malignancy and suppressive immune microenvironment of HCC, there are still high recurrence and metastasis rates even in HCC patients who have undergone radical resection. Therefore, it is of vital importance to conduct systematic surveillance of HCC recurrence, and there is an urgent need to precisely predict recurrence in patients with HCC. If tumour recurrence can be detected earlier, the survival and quality of life of HCC patients might be greatly improved.

Despite the rapid development in the treatment of HCC in recent decades, patients’ outcomes remain unsatisfactory. One of the reasons is that the early diagnosis system of HCC recurrence is not yet well developed, so our research team established a recurrence prediction model for HCC based on medical imaging such as computed tomography (CT) to predict HCC recurrence earlier, so that timely treatment measures can be taken.

We collected CT images from 537 clinical patients in two institutions and extracted valuable CT image features with 3D Slicer (v4.11, https://www.slicer.org). SPSS18.0 (SPSS Inc., Chicago, IL, United States) and R (version 4.0.3, https://www.rproject.org/) were used for statistical analyses and the prediction model of HCC recurrence was established jointly with AFP.

The radiomics scores calculated herein revealed significant differences between early and nonearly recurrence HCC patients. We combined radiomics and serum indicators to evaluate the risk of early recurrence, which was validated in an independent cohort. Our findings showed the value of predicting early recurrence in HCC patients by noninvasive indicators and might guide clinical decisions making. However, this is a retrospective analysis, and inherent biases are inevitable. We would conduct prospective trials with multimodality radiomics to confirm our results in future.

The preoperative radiomics model was shown to be effective for predicting early recurrence among patients with single HCC. Compared with pathological biopsy or other tests, this model is noninvasive and more convenient for HCC patients.

For HCC diagnosis, treatment, or prognosis assessment, more non-invasive methods are of great significance and needed.