A Multiparametric Fusion Radiomics Signature Based on Contrast-Enhanced MRI for Predicting Early Recurrence of Hepatocellular Carcinoma

Objectives The postoperative early recurrence (ER) rate of hepatocellular carcinoma (HCC) is 50%, and no highly reliable predictive tool has been developed yet. The aim of this study was to develop and validate a predictive model with radiomics analysis based on multiparametric magnetic resonance (MR) images to predict early recurrence of HCC. Methods In total, 302 patients (training dataset: n = 211; validation dataset: n = 91) with pathologically confirmed HCC who underwent preoperative MR imaging were enrolled in this study. Three-dimensional regions of interest of the entire lesion were accessed by manually drawing along the tumor margins on the multiple sequences of MR images. Least absolute shrinkage and selection operator Cox regression were then applied to select ER-related radiomics features and construct radiomics signatures. Univariate analysis and multivariate Cox regression analysis were used to identify the significant clinico-radiological factors and establish a clinico-radiological model. A predictive model of ER incorporating the fusion radiomics signature and clinico-radiological risk factors was constructed. The diagnostic performance and clinical utility of this model were measured by receiver-operating characteristic (ROC), calibration curve, and decision curve analyses. Results The fusion radiomics signature consisting of 6 radiomics features achieved good prediction performance (training dataset: AUC = 0.85, validation dataset: AUC = 0.79). The predictive model of ER integrating clinico-radiological risk factors and the fusion radiomics signature improved the prediction efficacy with AUCs of 0.91 and 0.87 in the training and validation datasets, respectively. Furthermore, the nomogram and ER risk stratification system based on the predictive model demonstrated encouraging predictions of the individualized risk of ER and gave three risk groups with low, intermediate, or high risk of ER. Conclusions The proposed predictive model incorporating clinico-radiological factors and the fusion radiomics signature derived from multiparametric MR images may be an effective tool for the individualized prediction of postoperative ER in patients with HCC.


Introduction
Hepatocellular carcinoma (HCC) is one of the most common malignant tumors and ranks as the fourth cause of cancerrelated death worldwide [1]. Surgical resection is the most effective curative treatment for HCC patients with wellpreserved liver function [2,3]. However, the clinical outcomes of patients after surgery are poor owing to the 5-year tumor recurrence rate reaching 70%. Mechanisms of tumour recurrence include true recurrence, which mainly manifests as early recurrence, and de novo tumours [4,5]. Early recurrence (ER) is defined as the detection of intrahepatic and/or extrahepatic HCC recurrence within the first 2 years after surgery. ER accounts for over 70% of tumor recurrence and is also relevant to a worse prognosis [6,7]. erefore, developing a highly reliable predictive tool is vital for surgical management, postoperative surveillance, and therapeutic interventions.
Some studies have shown that tumor histological factors, such as poor differentiation grade, microvascular invasion, and satellite nodules, are associated with ER, but these histological features cannot be accessed until postoperation [8][9][10][11][12][13]. Several HCC staging systems are helpful for prognostic evaluation, reasonable treatment option selection, and clinical research [1], such as the Barcelona Clinic Liver Cancer (BCLC) staging system, Tumor-Node-Metastasis (TNM) staging system, and the Japan Society of Hepatology (JSH) staging system. However, these systems can neither predict ER of HCC nor precisely provide quantitative risk measurements. Novel methods for predicting ER and acquiring more precise prognostic information are urgently needed.
Radiomics is a promising, noninvasive technique that can extract high-dimensional and quantitative imaging features from traditional medical images to provide information about tumor aggressiveness and heterogeneity and predict posttreatment survival of oncology [14,15]. Some studies on HCC have suggested that radiomics features are useful in predicting microvascular invasion, tumor recurrence, and treatment response [16][17][18]. To the best of our knowledge, few studies have evaluated the value of predicting postoperative ER of HCC with radiomics analysis based on multiparametric MR imaging to date. erefore, the purpose of our study was to develop and validate a predictive model of ER of HCC combining radiomics features and clinico-radiological risk factors based on multiparametric MR imaging. Meanwhile, the clinical values of the nomogram and ER risk stratification system based on the predictive model were also investigated.

Study Population.
is retrospective study was approved by our institutional ethics committee, and the requirement for informed consent was waived from each patient. e electronic medical records of our institution were reviewed for patients who underwent preoperative abdominal multiparametric MRI and had pathologically confirmed HCC between January 2015 and April 2018. e inclusion criteria of the patients were as follows: (a) multiparametric MR images with good image quality were available; (b) curative hepatectomy was performed within 1 month after MRI examination; and (c) clinico-radiological and follow-up information were complete. e exclusion criteria of the subjects were as follows: (a) the lesion was too small (<1 cm) to draw regions of interest (ROIs); (b) patients with antitumor treatments before surgery, such as repeat liver resection, local ablation, or transarterial chemoembolization; (c) postoperative adjuvant therapy was present; or (d) patients died of other diseases during the follow-up period.
Finally, a total of 302 patients (245 males and 57 females; median age, 5 years; and range, 23-86 years) were included in this study. is study consisted of 141 patients with ER and 161 patients without ER. ese patients were divided randomly into a training dataset (n � 211; 72 males and 39 females; median age, 57.2 ± 9.8 years; and range, 26-86 years) and a validation dataset (n � 91; 73 males and 18 females; median age, 57.6 ± 10.2 years; and range, 23-84 years) at a ratio of 7 : 3.
2.2. Follow-Up. All patients were followed up regularly after discharge. Serum α-fetoprotein (AFP) levels, liver function tests, and various imaging modalities (ultrasound, contrastenhanced CT, or MRI) of the abdomen were conducted to monitor recurrence of HCC during follow-up in the first month after liver resection and every 3 or 6 months thereafter [2,19]. Early recurrence was defined as intrahepatic and/or extrahepatic recurrence of HCC within 2 years after surgery. Recurrence-free survival (RFS) was defined as the interval between the date of surgery and the date of tumor recurrence.

MRI Protocol and Imaging
Analysis. A 1.5T MR scanner (Signa Excite HD, GE Healthcare, Milwaukee, WI, USA) with an 8-channel phased-array software coil was used to obtain abdominal multiparametric MR images. MRI sequences included axial in-phase and out-of-phase T1weighted imaging, axial T2-weighted imaging with fat suppression, and axis diffusion-weighted imaging (DWI). Axial dynamic contrast-enhanced fat-suppressed T1weighted images (3D Liver Acquisition with Volume Acceleration (3D-LAVA) sequence) were obtained in arterial phases (20-30 seconds), portal venous phases (60-70 seconds), and delayed phases (180 seconds) after bolus injection of 0.1 mmol/kg gadopentetate dimeglumine (Gd-DTPA, Magnevist, Bayer Schering, Berlin, Germany). e details of the MRI parameters are provided in Supplementary Table 1.

Clinico-Radiological Risk Factor Selection and Clinico-
Radiological Model Development. Routine clinical factors (age, sex, underlying liver disease, Child-Pugh class, histologic differentiation, status of microvascular invasion, etc.) of patients were collected from electronic medical records. e clinical factors and MR characteristics related to ER were selected. e t-test or Mann-Whitney U test was used to compare continuous variables, and the chi-squared test was used for categorical variables. Univariate analysis was performed to identify significant predictors of ER in the training dataset. en, those significant factors with a pvalue less than 0.05 were entered into the multivariate Cox regression analysis. Finally, the clinico-radiological predictive model was constructed in the training dataset with the significant factors of ER (Table 1).

Radiomics Workflow of MR Images.
e workflow of radiomics consists of tumor segmentation, radiomics feature extraction, radiomics feature selection, radiomics signature construction, model construction, and clinical utility ( Figure 1).

Tumor Segmentation and Radiomics Feature Extraction.
A radiologist with 4 years of work experience using ITK-SNAP software (version 3.8.0, https://www.itksnap. org/pmwiki/pmwiki.php?n�Main.HomePage) manually drew the region of interest (ROI) of the entire tumor on each transverse slice of T2-weighted images, diffusionweighted images (b-value is 800 s/mm 2 ), and arterial phase and portal venous phase images. Twenty tumors were randomly selected and then repeatedly segmented to remove unstable radiomics features, whose intraclass correlation coefficients (ICCs) were lower than 0.80. e Pyradiomics package (version 3.0.0, https://www. radiomics.io/pyradiomics.html) was used to extract 853 radiomics features (16 shape features, 19 first order features, 74 texture features, and 744 wavelet features) from each 3D segmentation. A total of 3412 radiomics features for each tumor (T2WI, DWI, and arterial phase and portal venous phase images) were obtained. e radiomics analysis was carried out by the guidelines of the Image Biomarker Standardisation Initiative [22].

Radiomics Feature Selection and Radiomics Signature Construction.
e values of radiomics features extracted from all sequences were normalized with z-scores. First, stable radiomics features with ICC values greater than 0.80 were utilized for further analysis (Supplementary Figure 1). Second, the least absolute shrinkage and selection operator was used to select statistically significant radiomics features (Supplementary Table 3, Supplementary Figure 2). ird, the radiomics signature of each MR sequence was constructed with multivariable Cox regression analysis [23]. e formulas of radiomics signatures are provided in the Supplementary Formula. e prediction performances of the radiomics signatures were compared. e final fusion radiomics signature was constructed using radiomics features from the MR sequences, which had a higher performance.

Predictive Model of ER Construction and Evaluation.
In the training dataset, the clinico-radiological risk factors and fusion radiomics signature were combined to develop a predictive model of ER with multivariable Cox regression analysis. e discriminative performance of the predictive model was quantitatively evaluated by the area under the curve (AUC) of the receiver operator characteristic (ROC) curve.
e accuracy, sensitivity, and specificity were also calculated. A nomogram was developed based on the predictive model of ER. Calibration curves were drawn to evaluate the calibration of the nomogram in the training dataset and validation dataset [24]. e net benefits under different threshold probabilities were quantified by decision curve analysis (DCA) to assess the clinical value of the predictive model [25].
2.9. ER Risk Stratification System. A risk stratification system of ER was constructed based on the total score of each individual, which was derived from the predictive nomogram. Its reliability was verified in the validation dataset.

Statistical Analysis.
All statistical analyses were performed with SPSS (version 23.0, Chicago, IL, USA) and R software (version 3.6.2, https://www.r-project.org) in this study. Detailed descriptions of the statistical methods and R packages are provided in Supplementary Method 1. A twotailed p < 0.05 was considered statistically significant.

Predictive Performances of Radiomics Signatures from Different MR Sequences.
e predictive performances of radiomics signatures using MR sequences are shown in Table 2. e radiomics signatures of arterial phase images and portal venous phase images had better predictive   e fusion radiomics signature had good predictive efficacy for ER (Table 2).

Predictive Model Construction and Evaluation.
A predictive model of ER that consisted of both clinicoradiological predictive factors and the fusion radiomics signature reached a satisfying performance. For the prediction of ER, the model displayed an AUC of 0.91 (95% CI: 0.87-0.95) with an accuracy, a sensitivity, and a specificity of 85.8%, 91.5%, and 78.7%, respectively, in the training dataset and an AUC of 0.87 (95% CI: 0.79-0.94) with an accuracy, a sensitivity, and a specificity of 81.3%, 88.6%, and 74.5%, respectively, in the validation dataset. e nomogram was developed based on the predictive model of ER, which is presented in Figure 3. e calibration curves of the nomogram (Figure 4) indicated that the predicted probabilities  A fusion radiomics signature was developed with arterial phase images and portal venous phase images. e predictive model consisted of a fusion radiomics signature and a clinico-radiological model. T2WI: T2-weighted imaging, DWI: diffusion-weighted imaging, SENS: sensitivity, SPEC: specificity, ACC: accuracy, and AUC: area under the curve. 6 Journal of Oncology were similar to the actual ER rate in the training and validation datasets. e decision curves showed that the predictive nomogram provided higher net benefits compared with the "treat-all" or "treat-none" scheme within the reasonable threshold probability ( Figure 5).

ER Risk Stratification
System. A risk stratification system of ER was also constructed on the basis of the aggregate score of each individual, which was derived from the predictive nomogram in the training dataset. All patients were divided into low-risk (156/302, 51.6%), intermediate-risk (73/302, 24.2%), and high-risk groups (73/302, 24.2%). In the training and validation datasets, ER rates were 14.9% and 21.4% for the low-risk group, respectively, while they were 59.6% and 57.7% for the intermediate-risk group and 98% and 100% for the high-risk group ( Figure 6). Median recurrence-free   Table 2).

Discussion
In this retrospective work, our research results show that the radiomics signatures can predict postoperative ER of HCC and the prediction model that combines clinico-radiological factors and the fusion radiomics signature has a better predictive efficacy for ER. In addition, the nomogram based on the suggested model shows a good prediction and discrimination ability and could be regarded as a noninvasive tool for the individualized prediction of ER. Furthermore, the risk stratification system developed in our study can successfully separate all of the patients into three distinct ER risk subgroups. In our study, some of the clinico-radiological factors, such as tumor margin, peritumoral enhancement, intratumor necrosis, microvascular invasion, and AFP, were not significantly correlated with ER of HCC in multivariable Cox regression analysis. ese results are inconsistent with those of previous studies [26][27][28]. e differences might be attributed to the bias of patient selection and different cut-off values of the laboratory factors applied in our and other studies.
ER of HCC occurs in 70% of patients and is also related to a worse prognosis [6,7]. erefore, developing a highly reliable predictive tool is important for surgical management, postoperative surveillance, and therapeutic interventions. Radiomics is a novel method that extracts many features from medical images and has been widely used to evaluate the aggressiveness and heterogeneity of HCC [29]. In this study, we developed a predictive model of ER in HCC based on multiparametric MR images with radiomics analysis. Six radiomics features were used to construct the fusion radiomics signature, including five features of arterial phase and one feature of portal venous phase. e performance of the fusion radiomics signature was similar to that of the arterial phase radiomic signature because the fusion radiomics signature was mostly derived from the five radiomics features of the arterial phase images. In addition, we also found that two radiomics features of the arterial phase images were associated with the heterogeneity of ER of HCC: dependence nonuniformity normalized (AP_wavelet.LHH_gldm_DNN) and small dependence emphasis (AP_wavelet.HLL_gldm_SDE). e dependence nonuniformity is usually used to measure the similarity of dependence throughout the image, and the small dependence emphasis is a measure of the distribution of the large dependencies, with a greater value indicative of less homogeneity. ese results indicate that radiomics features can offer more information on tumor biological behaviour and the tumor microenvironment, which are complementary to common imaging features. Furthermore, we integrated the fusion radiomics signature and the significant clinico-radiological factors to develop the combined predictive model of ER. e predictive model reached a satisfying performance. Meanwhile, our research results demonstrate that the nomogram and the risk stratification system based on the predictive model are effective tools for   Figure 6: Graphs represent the rates of early recurrence of HCC based on the three risk groups defined by the predictive model in the training and validation datasets.
the individualized prediction of ER risk and selection of optional treatment and surveillance regimens. For example, patients with a higher risk of ER may be undergoing liver transplantation instead of hepatectomy. If liver transplantation is impossible, appropriate adjuvant treatment and intensive surveillance should be considered [30,31]. ere are several limitations to the present study. One major limitation is that we performed a retrospective study in a single institution; thus, prospective and multicentric studies with considerably large datasets are needed for further validation of the robustness and reproducibility of the present prediction model. In addition, the MRI and radiomics features of the largest tumor were analysed when the lesions were multifocal in our study. erefore, the results of our study have a certain bias. ird, our study evaluated radiomics features extracted from DWI without apparent diffusion coefficient (ADC) values. e diagnostic performance for the predictive model can be further evaluated when combining the ADC value, one of the quantitative parameters, with other MRI and clinical data. Moreover, manual tumor segmentation was used in our present study, which could cause intra-and inter-user variabilities and can be time-consuming.
us, semiautomatic and automatic segmentation methods with high accuracy need to be developed.
In conclusion, the prediction model combining clinicoradiological factors and the fusion radiomics signature based on multiparametric MR images achieved an encouraging performance in predicting ER of HCC patients after curative hepatectomy. e nomogram and ER risk stratification system based on the model could have better potential for assisting in clinical decision-making and offering personalized therapies. Data Availability e data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
e authors declare that they have no conflicts of interest regarding the publication of this paper.