1. IntroductionHepatocellular carcinoma (HCC) is found to account for approximately 10% of cancer death worldwide, and it is particularly prevalent in Eastern and Southeastern Asian countries, including China [1]. The major risk factors of HCC are hepatic chronic diseases, especially the infection of hepatitis B virus and hepatitis C virus [2]. Despite improving prognosis by virtue of structured surveillance on known high-risk patients, HCC has very poor prognosis. Systemic therapy is conventionally used to treat patients with metastatic HCC and Sorafenib is the most commonly used first line treatment. However, it is not particularly well tolerated with treatment discontinuation due to side-effects in nearly 60% of patients [3]. Furthermore, as a small molecule multikinase inhibitor, both the VEGF receptor pathway and Raf kinase, critical to physiological function and homeostasis in many organs are blocked by Sorafenib. This can result in life-threatening complications, such as hemorrhage and cardiac events [4]. With locoregional therapies achieving excellent outcomes, it is paramount to accurately diagnose extra-hepatic disease in HCC patients.Although extrahepatic metastasis is commonly known to be associated with intrahepatic masses with high heterogeneity in computed tomography and vascular invasion, these qualitative factors are relatively subjective as for individual radiologists. Thus, the dual-tracer positron emission tomography (PET)/CT is an emerging modality that has shown much promise due to its high reported sensitivity and accuracy for metastatic disease, but its availability is limited [5,6,7]. Without a promising approach for indicating extrahepatic metastasis clinically, the decision to obtain uniform metastasis workup is most likely made by the treatment provider [8].While several biomarkers are reported to have a possible association with occurrence of metastasis of HCC, reliable biomarkers are yet to be standardized [9,10]. In past studies, quantitative image analysis is stated to be useful for deriving tumor dynamics in cellular and tissue level and developing imaging biomarkers, hence, could help to prognosticate underlying tumor biology [11,12]. At the same time, studies on radiomics of computed tomography (CT) of liver for predicting microvascular invasion (MVI), treatment outcome and recurrence of HCC are pointing to the possibility of radiomics in discovering imaging biomarkers related to HCC [12,13,14,15]. Prediction of MVI in 120 patients based on quantitative CT features yielded an area under the receiver-operating characteristic (ROC) curve (AUROC) of 0.80, positive predictive value of 63%, and negative predictive value of 85% [13]. A study proposed a radiomics Cox model for stratifying 88 patients with HCC treated with transarterial chemoembolization (TACE) into the high and low risk groups and compared its performance with the clinical score model [14]. Based on the proposed model, the hazard ratio (HR) of the high-risk group with reference to the low-risk group attained 7.42, which was higher than HR based on the clinical score model, 4.84 [14]. Early HCC recurrence prediction in 215 patients based on CT radiomics yielded an AUC of 0.817 (95% CI: 0.758–0.866), sensitivity of 0.794, and specificity of 0.699 [15]. The potential of imaging biomarkers in predicting extrahepatic metastasis of HCC could probably promote more efficient diagnosis and better treatment planning for the HCC patients.Radiomics is a developing field of image analysis that utilizes multiple data-mining algorithms to collect image features that are not visible to the naked eye and integrate them to obtain information for prediction or prognosis [11,16]. By constructing models accordingly, it could perform successful prediction and evaluation in certain clinical tasks [16]. Multiple studies have reported clear correlations between CT radiomics and clinical outcomes, while a further combination with selected clinical factors could achieve higher accuracies and clinical benefits [12,13,14,15,16]. Despite the potential, the generalization of radiomics as a clinical indicator still requires numerous refinements and standardizations to allow clinicians to confidently implement radiomics in patient management [11,17]. From previous studies, CT radiomics has shown the potential to predict MVI in HCC and recurrence of HCC [13,15], whereas the metastatic rate was predicted by clinical features only [5,8]. As far as we know, no published study has assessed the association between CT radiomics of HCC tumor and its metastatic risk. A pilot study would provide ground for deeper exploration in this aspect.Contrast-enhanced CT is preliminarily used for diagnosis and staging of HCC, which is currently irreplaceable standard-of-care modality in HCC management [18]. If the radiomic features related to extrahepatic metastasis could be mined from the CT images of liver, additional information can be obtained without extra radiation exposure on the potential metastatic sites. The purpose of this study is to investigate the possibility of using radiomics features, obtained from contrast-enhanced CT images of the liver by a computational approach, to identify extrahepatic metastasis in HCC patients. 4. Discussion

In the present study, we explored the possibility of using radiomics in contrast-enhanced CT to be a predictive indicator for metastasis disease in HCC patients. A radiomic model was constructed, and it showed its potential to individually identify HCC patients with high likelihood to have extrahepatic metastasis.

CT Imaging has become a crucial imaging modality in the management of HCC [18]. In recent years, the application of radiomics has allowed researchers to mine clinical and prognostic information from medical images by quantifying the phenotypic characteristics of tumors [16,29]. Various studies showed that CT images could predict the prognosis of HCC patients [12,13,14,15,16]. Detection of extrahepatic metastasis allows physicians to provide appropriate treatments for HCC patients although no previous study has explored the use of radiomics [6]. Thus, we designed this study to investigate possible predictors of extrahepatic metastasis, an important factor for patient prognosis and survival [5].Radiomic features of different categories can quantify distinct intratumoral characteristics and thus reflect tumor complexity in multiple aspects. Despite the large number of features being tested, we further performed binary logistic regression and selected the first eight features of higher reproducibility and stability to avoid possible over-fitting of our model [25,26]. Half of the eight selected radiomic features were GLSZM based, one of them was shape based, one of them was GLDM based, and two of them were of first-order category.GLSZM based features measure the spatial interrelationship of adjacent groups of grey level voxels in 13 directions three-dimensionally [30]. Four GLSZM features relevant to the nonuniformity of the grey level of the tumor in CT images were identified, indicating that the tumor heterogeneity was closely related to the possibility of metastasis. With generally higher magnitudes in GLSZM features of metastatic cases, the result can be related to the finding that textural heterogeneity in tumors could probably indicate metastasis, and hence poor prognosis and survival [12,31]. GLSZM based features have an advantage in that they are relatively more reproducible regardless of the segment accuracy and the interobserver reliability. Less precise segmentation could still generate similar results as the heterogeneity is often more significant in the center of the tumor but more subtle on the edges of ROIs [32].Shape features quantify the shape and size of the ROI, including diameter, surface area and irregularity [30]. The selected shape feature measures the maximum axial diameter of the HCC drawn. Similar to the findings of Natsuizaka et al. [6], our results show that the longer the mean tumor diameter, the more likely the patient belonged to the metastatic group (p = 0.007).GLDM based features mathematically describe the distributions of different grey levels within the ROI [33]. The small dependence low grey level emphasis measures the magnitude of low grey level distribution and indicates the density of voxels with low grey value in the ROI. We found that a smaller distribution of low grey value voxels may indicate higher likelihood of metastasis. This finding was consistent with the study by Mao, et al. [34], who found that less distribution of low grey level in ROI of arterial phase CT could be correlated to high-grade HCC, as it might reflect higher contrast enhancement and vascularity. High vascularity of HCC often promotes faster growth, infiltration, and invasion, thus increasing the likelihood of extrahepatic metastasis [5].First-order features quantify the histogram distribution of the intensity values of the voxels in the ROI [35]. The two identified first-order features indicated that a histogram with higher total energy and maximum could stipulate extrahepatic metastasis. Kim et al. [14] reported a similar relationship between high energy in histogram and HCC tumor heterogeneity which could be related to metastasis, while a study by Peng, et al. [36] reported that a higher maximum in histogram could indicate microvascular invasion which directly increases the risk of extrahepatic metastasis. The first-order features we identified agreed with those in previous research.On the external validation, the performance metrics of logistic regression were all comparable or better than SVM and VGG16. Significant difference in specificity and AUC between logistic regression and VGG16 was identified. VGG16 performed poorly because the relatively small training set was inadequate to train a very large network with huge number of weights. Although no significant difference between logistic regression and SVM was identified, the logistic regression yielded a more meaningful model where the coefficients represent the change in log odds of metastasis per unit change in the corresponding radiomic features. Based on logistic regression, the resultant radiomic model had AUCs of 0.914, 0.944, and 0.744 on the training, test and external validation sets respectively, which was comparable if not better than the performance of various similar radiomic models established by other researchers for predicting pathological or surgical outcomes of HCCs (AUCs: 0.670–0.859) [13,15,34,36,37,38]. The innovation of this study is that the radiomic model based on the image information of tumor region only can stratify the HCC patients into risk groups of extrahepatic metastases and support the decision for metastasis workups.In the present study, we also identified some clinical features that might also have the capability to predict extrahepatic metastasis of HCC, including tumor diameter and number of lesions. While the clinical significance of tumor diameter was stated by various studies and was reflected in our radiomics model, the predictive power of number of lesions is controversial [5,6]. We performed univariate analysis on the HCC lesion numbers of cases and ranked it with the radiomic features extracted. The number of lesions was found to have less significant effect on extrahepatic metastasis when compared to the radiomic features. We also built a second model by combing the 8 selected radiomic features and the number of lesions and tested it using the same test set. The accuracy of the second model was not superior to our original model (accuracy: 75.0% vs. 83.3%). While our findings suggest that tumor numbers have limited predictive power for extrahepatic metastasis, studies by Uchino et al. [5] and by Natsuizaka et al. [6] reported it as an essential indicator for HCC metastasis in clinical practice. These contradictory conclusions might be resulted by various reasons. First, with a limited number of samples in our study, we might be unable to fully stratify metastatic and non-metastatic patients by a single clinical factor. Moreover, the number of tiny satellite lesions may not be completely reported in the radiologist reports, badly affecting the representing power of tumor numbers in our analysis. It is undeniable that number of lesions is a clinical feature that is far more accessible to the clinicians when compared to radiomic features, which might also be a reason for the tumor number to be a prognostic indicator for extrahepatic metastasis in hospital settings.Our study has some limitations. The analytical results might be subjected to different standards in image acquisition, postprocessing and reconstruction across centers. Batch harmonization techniques, such as global scaling and z-standardization, were proposed to minimize feature variabilities [39]. A thorough assessment of the most appropriate technique is required for developing a radiomic model involving multiple centers. The study was also limited by a small sample size that could lead to instability in extraction and analysis of radiomic features, while the imbalanced data set might also cause inaccuracies in feature selection and analysis, although it has undergone SMOTE. Future studies with more comprehensive and larger samples are required to further verify our findings. We only extracted radiomic features from the largest HCC lesion in each case, as there were satellite lesions that were difficult to draw and might be subjected to measurement error [40]. The CT images were acquired by several different CT scanners over a few years of time. Differences between CT scanners, a change in protocols, use of different contrasts, and evolved reconstruction and postprocessing techniques might affect the radiomic features. Although the effects could be unintentionally reflecting the clinical reality that multiple CT scanners and protocols might be used clinically, it is still one of the limitations of our research design [14,40]. Since additional information, such as histological features, were not quite considered in the present study, future studies are needed to further interpret the radiomic features with biological markers. We believe that the modification of the model into a cluster-based search algorithm will allow clinicians to retrieve cases with similar radiomics features and clinical metastatic factors. Then, the model can assist clinicians in determining the MET possibility of a newly registered HCC case and suggesting which organs are at a higher MET risk.