Construction and application value of CT-based radiomics model for predicting recurrence of early-stage hepatocellular carcinoma after resection
10.3760/cma.j.issn.1673-9752.2020.02.014
- VernacularTitle:基于CT检查影像组学早期肝细胞癌切除术后肿瘤复发的预测模型构建及其应用价值
- Author:
Guwei JI
1
;
Ke WANG
;
Xiaofeng WU
;
Yongxiang XIA
;
Changxian LI
;
Hui ZHANG
;
Hongwei WANG
;
Mingyu WU
;
Bing CAI
;
Xiangcheng LI
;
Xuehao WANG
Author Information
1. 南京医科大学第一附属医院肝胆中心 210029
- From:
Chinese Journal of Digestive Surgery
2020;19(2):204-216
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct a computed tomography (CT)-based radiomics model for predicting tumor recurrence of early-stage hepatocellular carcinoma (HCC) after resection, and explore its application value.Methods:The retrospective cohort study was conducted. The clinicopathological data of 243 patients with early-stage HCC who underwent hepatectomy in 2 medical centers between January 2009 and December 2016 were collected, including 165 in the First Affiliated Hospital of Nanjing Medical University and 78 in the Wuxi People′s Hospital. There were 182 males and 61 females, aged from 30 to 86 years, with a median age of 57 years. According to the random numbers showed in the computer, 243 patients were randomly assigned into training dataset consisting of 162 patients and test dataset consisting of 81 patients, with a ratio of 2∶1. Using radiomics technique, a total of 3 384 radiomics features were extracted from the tumor and its periphery at arterial-phase and portal-phase images of CT scan. In the training dataset, a radiomics signature was constructed and predicted its performance after dimension reduction of stable features by using aggregated feature selection algorithms [feature ranking via maximal relevance and minimal redundancy (MRMR) combined with random survival forest (RSF) + LASSO-COX regression analysis]. Risk factors for tumor recurrence were selected using the univariate COX regression analysis, and two radiomics models including radiomics 1 (preoperative) and radiomics 2 (postoperative) were constructed and predicted their performance using backward stepwise multivariate COX regression analysis. The two models were validated in the training and test dataset. Observation indicators: (1) follow-up; (2) construction of HCC recurrence-related radiomics signature for early-stage HCC after resection; (3) prediction performance of HCC recurrence-related radiomics signature for early-stage HCC after resection; (4) construction of HCC recurrence-related radiomics prediction model for early-stage HCC after resection; (5) validation of HCC recurrence-related radiomics prediction model for early-stage HCC after resection; (6) comparison of the prediction performance of radiomics model with that of other clinical statistical models and current HCC staging systems; (7) stratification analysis of postoperative recurrence risk based on radiomics models for early-stage HCC after resection. Patients were followed up using outpatient examination or telephone interview once every 3 months within the first 2 years and once every 6 months after 2 years. The follow-up included collection of medical history, laboratory examination, and abdominal ultrasound examination. Contrast-enhanced CT or magnetic resonance imaging (MRI) examination was performed once every 6 months, and they were performed in advance on patients who had suspected recurrence based on laboratory examination or abdominal ultrasound for further diagnosis. Follow-up was up to January 2019. The endpoint was time to recurrence, which was from the date of surgery to the date of first detected disease recurrence or metastasis. Measurement data with normal distribution were represented as Mean± SD, and comparison between groups was analyzed by the t test. Measurement data with skewed distribution were described as M (range), and comparison between groups was analyzed by the Mann-Whitney U test. Count data were described as absolute numbesr or percentages, and comparison between groups was analyzed using the chi-square test. The survival curve and survival rate were respectively drawn and calculated by the Kaplan-Meier method, and the survival analysis was performed using the Log-rank test. Serum alpha-fetoprotein level was analyzed after the natural logarithm transformation. X-tile software was used to select the optimal cut-point for continuous markers. Results:(1) Follow-up: all the 243 HCC patients received follow-up. Patients in the training dataset were followed up for 4.2-109.2 months, with a median follow-up time of 51.6 months. Patients in the test dataset were followed up for 12.7-107.6 months, with a median follow-up time of 73.2 months. The 2-, 5-year disease-free survival rates were 77.8% and 53.1% of the training dataset respectively, versus 86.4% and 61.7% of the test dataset. There was no significant difference in terms of disease-free survival between two datasets ( χ2=1.773, P>0.05). (2) Construction of HCC recurrence-related radiomics signature for early-stage HCC after resection: of the 3 384 radiomics features, 2 426 radiomics features with high stability were selected for analysis. There were 37 radiomics features identified after combining the top 20 radiomics features ranked by MRMR and RSF algorithms. LASSO-COX regression algorithm further reduced their dimensionality to retain 7 radiomics features and construct a radiomics signature. The indicators including region, scanning phase, and weighting coefficient of above mentioned seven features were Feature 1 (peritumoral, arterial phase, 0.041), Feature 2 (peritumoral, arterial phase, -0.103), Feature 3 (peritumoral, arterial phase, -0.259), Feature 4 (intratumoral, arterial phase, 0.211), Feature 5 (peritumoral, portal venous phase, -0.170), Feature 6 (intratumoral, portal venous phase, 0.130), and Feature 7 (intratumoral, portal venous phase, 0.090), respectively. Radiomics signature score=0.041×Feature 1-0.103×Feature 2-0.259×Feature 3+ 0.211×Feature 4-0.170×Feature 5+ 0.130×Feature 6+ 0.090×Feature 7. (3) Prediction performance of HCC recurrence-related radiomics signature for early-stage HCC after resection: the radiomics signature showed favorable prediction performance in both training and test datasets, with respective C-index of 0.648 [95% confidence interval ( CI): 0.583-0.713] and 0.669 (95% CI: 0.587-0.750). (4) Construction of HCC recurrence-related radiomics prediction model for early-stage HCC after resection: results of univariate analysis showed that ln(serum alpha-fetoprotein), liver cirrhosis, tumor margin status, arterial peritumoral enhancement, intratumoral necrosis, radiomics signature, satellite nodules, and microvascular invasion were related factors for tumor recurrence after resection of early-stage HCC ( hazard ratio=1.202, 1.776, 1.889, 2.957, 1.713, 4.237, 4.364, 4.258, 95% CI: 1.083-1.333, 1.068-2.953, 1.181-3.024, 1.462-5.981, 1.076-2.728, 2.593-6.923, 2.468-7.717, 2.427-7.468, P<0.05 ). Results of multivariate analysis showed that the radiomics model 1 (preoperative) consisted of ln(serum alpha-fetoprotein), tumor margin status, and radiomics signature ( hazard ratio=1.145, 1.838, 3.525, 95% CI: 1.029-1.273, 1.143-2.955, 2.172-5.720, P<0.05); the radiomics model 2 (postoperative) consisted of ln(serum alpha-fetoprotein), radiomics signature, microvascular invasion, and satellite nodules ( hazard ratio=1.123, 2.386, 3.456, 3.481, 95% CI: 1.005-1.254, 1.501-3.795, 1.863-6.410, 1.891-6.408, P<0.05). Risk prediction formulas: radiomics model 1 = 0.135×ln(serum alpha-fetoprotein)+ 0.608×tumor margin status (0: smooth; 1: non-smooth)+ 1.260×radiomics signature; radiomics model 2 = 0.116×ln(serum alpha-fetoprotein)+ 0.870×radiomics signature + 1.240×microvascular invasion (0: absent; 1: present)+ 1.247×satellite nodules (0: absent; 1: present). (5) Validation of HCC recurrence-related radiomics prediction model for early-stage HCC after resection: in both training and test datasets, radiomics model 1 provided good prediction performance, with respective C-index of 0.716 (95% CI: 0.662-0.770) and 0.724 (95% CI: 0.642-0.806), while radiomics model 2 provided better prediction performance, with respective C-index of 0.765 (95% CI: 0.712-0.818) and 0.741 (95% CI: 0.662-0.820). Calibration curves demonstrated good agreement between model-predicted probabilities and observed outcomes. (6) Comparison of the prediction performance of radiomics model with that of other clinical statistical models and current HCC staging systems: in the training dataset, the prediction performance of radiomics model 1 for tumor recurrence after resection of early-stage HCC was significantly different from that of ERASL model (preoperative), Barcelona clinic liver cancer (BCLC) staging, Hong Kong liver cancer (HKLC) staging, and cancer of the liver Italian program (CLIP) classification (C-index=0.562, 0.484, 0.520, 0.622, 95% CI: 0.490-0.634, 0.311-0.658, 0.301-0.740, 0.509-0.736, P<0.05); the prediction performance of radiomics model 2 for tumor recurrence after resection of early-stage HCC was significantly different from that of ERASL model (postoperative), Korean model, and the eighth edition TNM staging (C-index=0.601, 0.523, 0.513, 95% CI: 0.524-0.677, 0.449-0.596, 0.273-0.753, P<0.05). In the test dataset, the prediction performance of radiomics model 1 for tumor recurrence after resection of early-stage HCC was significantly different from that of ERASL model (preoperative), BCLC staging, HKLC staging, CLIP classification (C-index=0.540, 0.473, 0.504, 0.545, 95% CI: 0.442-0.638, 0.252-0.693, 0.252-0.757, 0.361-0.730, P<0.05); the prediction performance of radiomics model 2 for tumor recurrence after resection of early-stage HCC was significantly different from that of ERASL model (postoperative), Korean model, and the eighth edition TNM staging (C-index=0.562, 0.513, 0.521, 95% CI: 0.451-0.672, 0.399-0.626, 0.251-0.791, P<0.05). (7) Stratification analysis of postoperative recurrence risk based on radiomics models for tumor recurrence after resection of early-stage HCC: according to the analysis of X-tile, the score of radiomics model 1 < 1.4 (corresponding to total points < 62.0 in nomogram) was classified into low-risk group while the score of radiomics model 1 ≥ 1.4 (corresponding to total points ≥ 62.0 in nomogram) was classified into high-risk group. The score of radiomics model 2 < 1.7 (corresponding to total points < 88.0 in nomogram) was classified into low-risk group while the score of radiomics model 2 ≥ 1.7 (corresponding to total points ≥ 88.0 in nomogram) was classified into high-risk group. In the training dataset, the 2- and 5-year recurrence rates were 14.1%, 35.3% for low-risk patients and 63.0%, 100.0% for high-risk patients, which were predicted by radiomics model 1. There were significant differences between the two groups ( χ2= 70.381, P<0.05). The 2- and 5-year recurrence rates were 12.9%, 38.2% for low-risk patients and 81.8%, 100.0% for high-risk patients, which were predicted by radiomics model 2. There were significant differences between the two groups ( χ2= 98.613, P<0.05). In the test dataset, the 2- and 5-year recurrence rates were 5.6%, 29.3% for low-risk patients and 70.0%, 100.0% for high-risk patients, which were predicted by radiomics model 1. There were significant differences between the two groups ( χ2= 64.453, P<0.05). Ther 2- and 5-year recurrence rates were 5.7%, 28.1% for low-risk patients and 63.6%, 100.0% for high-risk patients, which were predicted by radiomics model 2. There were significant differences between the two groups ( χ2= 58.032, P<0.05). Conclusions:The 7-feature-based radiomics signature is built by selection of CT radiomics features in this study, and then HCC recurrence-related radiomics prediction model for early-stage HCC after resection is constructed. The proposed radiomics models can complement the existing clinical-radiological-pathological prognostic sources, accurately and individually predict tumor recurrence risk preoperatively and postoperatively, which facilitate clinical decision-support for patients with early-stage HCC.
