Jn-W W b c # S-Ru Fu # J Zn # Don-Sn Gu b c # Xo-Qun L Xu-Don Cn Su-Ton Zn b c Xo-F H Jn-Fn Yn L-Gon Lu J Tn b c j k

a Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China

b Beijing Key Laboratory of Molecular Imaging, Beijing 100190, China

c University of Chinese Academy of Sciences, Beijing 10 0 049, China

d Zhuhai Interventional Medical Center, Zhuhai Precision Medical Center, Zhuhai People’s Hospital, Zhuhai Hospital of Jinan University, Zhuhai 5190 0 0, China

e Department of Radiology, Zhuhai Precision Medical Center, Zhuhai People’s Hospital, Zhuhai Hospital of Jinan University, Zhuhai 5190 0 0, China

f Department of Interventional Treatment, Zhongshan City People’s Hospital, Zhongshan 528400, China

g Department of Radiology, Shenzhen People’s Hospital, Shenzhen 5180 0 0, China

h Interventional Diagnosis and Treatment Department, Nanfang Hospital, Southern Medical University, Guangzhou, 510 0 0 0, China

i Department of Radiology, Yangjiang People’s Hospital, Yangjiang 529500, China

j Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Medicine and Engineering, Beihang University, Beijing 100191, China

k Engineering Research Center of Molecular and Neuro Imaging of Ministry of Education, School of Life Science and Technology, Xidian University, Xi’an 710126, China

Keywords:Hepatocellular carcinoma Macrovascular invasion Radiomics Computed tomography Prognosis

ABSTRACT Background: Macrovascular invasion (MaVI) occurs in nearly half of hepatocellular carcinoma (HCC) patients at diagnosis or during follow-up, which causes severe disease deterioration, and limits the possibility of surgical approaches.This study aimed to investigate whether computed tomography (CT)-based radiomics analysis could help predict development of MaVI in HCC.

Introduction

Hepatocellular carcinoma (HCC) is the sixth most common malignancy and the fourth leading cause of cancer-related death [1].Macrovascular invasion (MaVI) occurs in 44%-62% of HCC patients at diagnosis or during follow-up, which is a great challenge for HCC management [ 2 , 3 ].Development of MaVI causes severe disease deterioration, and limits the possibility of surgical approaches,or even locoregional treatments [3].Once occurs, the expected median survival time is only 6–8 months [4].

Recent studies proved that aggressive treatments such as surgical resection, transcatheter arterial chemoembolization (TACE), and external beam radiotherapy would provide survival benefits [ 2 , 5 , 6 ].However, the application of these treatments is restricted to condition that MaVI does not cause liver dysfunction [ 7 , 8 ].Thus, there is an essential need to identify patients at high risk of MaVI development to perform systematic monitoring and offer timely intervention.To date, reported cases were mainly focused on microvascular invasion (MiVI) prediction, and limited studies were systematically conducted on MaVI subsequent occurrence prediction [9–12].Effi-cient and early prediction of highly aggressive MaVI development in HCC remains unmet.

Considering the role of noninvasive imaging in HCC diagnosis and treatment, prediction of MaVI through a radiological method is rational and promising [13].As a widely-applied technique, radiomics would serve as an effective tool to address the early prediction of MaVI outcomes.Beyond traditional assessment of morphological characteristics, radiomics captures textural or granular changes on preoperative radiological images that may depict tumor heterogeneity, as well as pathophysiology [14].By integrating key imaging features and clinical factors, this machine learningbased technique could automatically recognize archetypal subtypes related to targeted clinical issues [15].Radiomics can be successfully applied to predict lymph node metastasis in colorectal cancer,pathologic complete response in breast cancer, and relevant genetic mutation in lung cancer [16–18].In terms of HCC, radiomics could achieve prediction of MiVI and survival outcomes, which highlighted its effectiveness in HCC management [19].

Therefore, in this multicenter study, we aimed to develop and validate a predictive model integrating both quantitative radiomics signature and clinical factors to noninvasively identify patients at high risk of MaVI development.Meanwhile, the validity of radiomic features is assessed to predict the time of MaVI development, providing a reliable basis for personalized treatment planning in HCC management.

Methods

Patients

In this multicenter study, patients were collected from 5 Chinese hospitals.Chinese Clinical Trial Registry identifier of the study was ChiCTR180 0 016165.We totally included 226 HCC patients with MaVI and prognosis follow-ups from April 2007 to November 2016.The study was approved by the Ethics Committee of Zhuhai People’s Hospital (2018042601).Informed consent was waived for this retrospective study.All patient records and information were anonymized and de-identified prior to analysis.

The inclusion criteria were as follows: 1) HCC diagnosed by pathological or clinical confirmation; 2) receipt of preoperative computed tomography (CT) at diagnosis; and 3) developed MaVI confirmed by radiological approach during regular follow-up, death occurred and without MaVI development, or without disease progression and MaVI development for more than three years.The exclusion criteria were as follows: 1) MaVI occurred at diagnosis; 2)receipt of initial treatments other than standard therapies; and 3)irregular follow-up.The patient recruitment pathway is shown in Fig.1.

Treatments and follow-ups

Patients were initially treated by the guideline-recommended therapies, including liver resection, TACE, and ablation [ 3 , 5 ].Treatment strategy was decided by a multi-discipline team with consideration of factors including tumor characteristics, liver function,and patients’ will.

The interval between follow-ups was 4–6 weeks.Chest X-ray,abdominal CT/magnetic resonance imaging (MRI), and necessary laboratory tests were performed according to the guidelines during the postoperative check [ 3 , 4 ].Specifically, when extrahepatic metastasis was suspected, an additional CT or MRI for the suspected region was acquired.

Outcomes

MaVI presence was determined by radiological approach based on 4-phase dynamic CT with the following criteria: an intraluminal filling defect adjacent to the primary tumor in a portal and/or hepatic vein, an enhancement of the filling defect on the arterial phase, and a washout on the portal or delayed phases [ 5 , 20 ].The confirmation of MaVI was decided by two radiologists with 10 years’ working experience and censored by a senior radiologist with 20 years’ working experience.

Primary outcome was the time of MaVI development, which was defined as the interval from initial HCC diagnosis to the date when MaVI was confirmed radiologically.Secondary outcomes included progression-free survival (PFS) and overall survival (OS).PFS was defined as the time from initial diagnosis to the date of disease progression or death.OS was defined as the time from initial treatment to the date of death.

Image data acquisition

The parameters of CT scanning and contrast agency injection in each collaborative hospital are listed in Table S1.Given that HCC had a clearer boundary in the portal phase than other phases, we used CT images on portal phase as the analytical imaging data.

Candidate clinical factors and radiological characteristics

Clinical variables are reported in Table 1.Radiological characteristics included tumor number, diameter, location (left lobe/right lobe/cross lobe/both lobes), shape (noninvasive/invasive), corona sign with lowattenuation, mosaic architecture, node-in-node sign,and tumor enhancement (<25%/25%–50%/>50%–75%/>75%).Examples of detailed qualitative radiological characteristics are shown in Fig.S1.

Table 1Baseline characteristics in training and validation cohorts.

Table 2Predictive performance of the clinical model, radiomics signature, and CRIM.

Predictive model construction

Model development process was structured in 4 phases: imaging quantification, radiomics signature construction, predictive clinical factors selection, and integrated model construction.The flowchart is shown in Fig.2.Patients were divided into a training cohort (two hospitals,n= 154) and an external validation cohort(three hospitals,n= 72) by a center-independent split.The training cohort was used for model construction.Model assessment was conducted in both training and external validation cohorts.

Fig.1.Patient recruitment pathway.HCC: hepatocellular carcinoma; CT: computed tomography; PD: progression disease.

Fig.2.Model development flowchart.RFE: recursive feature elimination; RF: random forest; MaVI: macrovascular invasion; ROC: receiver operating characteristics; AUC:area under the curve; CRIM: clinical-radiomics integrated model; PFS: progression-free survival.

Imaging quantification

A total of 1217 radiomic features were automatically extracted on a manual-segmented tumor lesion which was delineated by two radiologists in a blinded fashion.A detailed description of tumor segmentation is provided in Supplementary Appendix E1.The radiomic features could be classified as 5 types: shape and size, firstorder statistics, texture, wavelet, and our newly designed –peel-off features.Moreover, we quantified four semantic features that were based on empirical experience of radiologists, including tumor border clarity, tumor circularity, number of nodule heaves on the border, and difference between center and cavity.Detailed definition and description of radiomic features are shown in Supplementary Appendix E2.

Regarding feature robustness analysis, feature stability and reproducibility were examined by concordance correlation coeffi-cients (CCCs) and intra-class correlation coefficients (ICCs), respectively.We randomly chose 20 patients to accept multisegmentations, including multiple clinician segmentation, multiple time-point segmentation, and perturbation segmentation.A detailed description of the multi-segmentation process is provided in Supplementary Appendix E3.

Radiomics signature construction

Radiomics signature construction consists of two processes: feature selection and prediction modeling.Feature selection is designed to select features to correlate with MaVI to the maximum extent.Redundant and irrelevant features are removed.The aim of prediction modeling is to choose the befitting machine learning classifier based on selected features for MaVI occurrence status prediction.Concerning modeling methodology, we conducted 16 most commonly used feature selection algorithms and three classifiers.Detailed description of the algorithms is shown in Supplementary Appendix E4.Referring to the optimal classifier, after statistics, three most commonly used algorithms in radiomics studies were found to be logistic regression (LR), random forest (RF),and support vector machine (SVM).In total, 48 modeling strategieswere implemented and compared.Recursive feature elimination(RFE) and RF were adopted as the final feature selection and prediction modeling algorithms, respectively.All algorithm implementations were performed in Python with PyCharm version 17.02.1.

Selection of predictive clinical factors

Univariable analysis was initially performed to select clinical factors correlated with MaVI development withPvalue less than 0.1.Furthermore, multivariable analysis with forward selection likelihood ratio was applied with aPvalue less than 0.05.Qualified clinical factors were treated as candidate parameters for the following integrated model construction.Moreover, a clinical model was built only based on clinical parameters by RF modeling.

Integrated model construction

We combined the clinical candidate parameters along with the selected radiomic features by RF modeling to construct the gathered network - CRIM.Heatmap was plotted to represent the correlation between the clinical factors and the selected radiomic features.

Time of MaVI development prediction

Multivariable Cox regression modeling was used to determine the independent risk factors in patients who underwent subsequent MaVI development (n= 48).Kaplan-Meier analysis with logrank test was used to stratify the high-risk and low-risk groups depending on MaVI development, PFS, and OS.

Model assessment

We adopted receiver operating characteristic (ROC) curves, area under the curve (AUC), accuracy, and confusion matrix as performance indicators.Comparisons between ROC curves were evaluated by the DeLong’s test.The Cox-regression model was assessed by C-index.Decision curve analysis was used to explore the clinical validity.

To assess the generalization of the developed models, we evaluated the performance in each independent external validation hospital that acquired CT images with different scanning parameters.Furthermore, we performed a stratification analysis according to sex, Barcelona Clinic Liver Cancer (BCLC) staging, and alphafetoprotein (AFP), as well as treatment manners.

Statistical analysis

Differences between the training and external validation cohorts, and between the MaVI and non-MaVI groups in terms of demographic and clinical variables were assessed with Pearson’s Chisquare tests for categorical variables, and Mann-WhitneyUtest for continuous variables.Statistical analysis was performed with SPSS version 18.0 (SPSS Inc., Chicago, IL, USA) and R software, Version 3.4.1 ( www.R-project.org ).Statistical significance was defined with a two-sidedPvalue less than 0.05.

Results

Baseline characteristics

The baseline characteristics are summarized in Table 1.There were no significant differences between the training and external validation cohorts in either demographic, clinical, or radiological characteristics, except for Child-Pugh class (P= 0.015) and hepatitis (P= 0.047).

During the follow-up, 48 (21.2%) of patients were found with subsequent MaVI development and 178 (78.8%) of patients without MaVI occurrence before death or the end of follow-up.There was no significant difference in the distribution of subsequent MaVI occurrence between the training and external validation cohorts (P= 0.805).Baseline characteristics of these 48 patients with MaVI development is provided in Table S2.BCLC staging and tumor number were significantly different between the MaVI and non-MaVI groups.

The median time of MaVI development was 6.9 months.In the MaVI group, the median PFS and OS were 5.7 months and 15.3 months, respectively.In the non-MaVI group, the median PFS and OS were 10.8 months and 26.8 months, respectively.

MaVI-related clinical and radiological factors

Node-in-node sign [odds ratio (OR): 3.22,P= 0.007]was selected as the only risk factor.The AUC of the clinical model were 0.638 and 0.487 in the training and external validation cohorts, respectively.Detailed predictive indicators of the clinical model are shown in Table 2.Confusion matrices of the clinical model are shown in Fig.S2 A, B.The violin graph of the clinical model and ROC curves are shown in Fig.3 A-C, respectively.

Radiomics signature

A set of 485 features eventually went into a robust feature pool for the radiomics analysis.The robust radiomic feature distribution per stability and reproducibility is shown in Fig.S3.

Fourteen radiomic features were selected to construct the radiomics signature.The selected feature name, formula, definition,and AUC value are shown in Table S3.The radiomics signature was distributed with a significant difference between the MaVI and non-MaVI groups in both the training and external validation cohorts (P<0.001).The AUCs of the radiomics signature were 0.965 in the training cohort and 0.957 in the external validation cohort.Detailed predictive indicators of the radiomics signature are shown in Table 2.Confusion matrices of the radiomics signature in the training and external validation cohorts are shown in Fig.S2 C, D.The violin graph of the radiomics signature and ROC curves are shown in Fig.3 A, B, D, E, respectively.The predictive outcomes of the other 47 radiomics modeling strategies are shown in Table S4.

CRIM model

The final integrated model –CRIM showed optimal ability to predict MaVI development with AUCs of 0.986 and 0.979 inthe training and external validation cohorts, respectively.Detailed predictive indicators of CRIM are shown in Table 2.CRIM also achieved extraordinary performance with AUCs of 0.956, 1.0 0 0, and 1.0 0 0 in external validation cohorts of Zhuhai People’s Hospital,Shenzhen People’s Hospital, and Nanfang Hospital (Table S4), respectively.Detailed results of independent external validation are shown in Table S5.Confusion matrices of CRIM in training, combined external validation and each independent external validation cohort are shown in Fig.S2E-I.The violin graph of the integrated model and ROC curves are shown in Fig.3 A, B, E, respectively.

CRIM presented with superior performance than the clinical model in both training and external validation cohorts (P<0.001),which highlighted the validity of the radiomics signature.The heatmap manifesting the correlation among clinical factor (nodeto-node sign) and radiomic features is shown in Fig.S4.

Stratification analysis

In sex subgroups, the AUC achieved 0.985/0.977 (training/validation) in males and 0.978/0.981 (training/validation) in females.In BCLC staging HCCs, CRIM also showed high AUCs of 0.992/0.985 (training/validation) in the early stage, and 0.977/0.994(training/validation) in the advanced stage.Regarding AFP, AUCs were 0.989/0.972 (training/validation) in the subgroup with AFP less than 25 ng/mL; 0.983/1.0 0 0 (training/validation) with AFP between 25 and 100 ng/mL; and 0.934/0.985 (training/validation)with AFP over 100 ng/mL.Detailed stratification results are shown in Table S6.

In particular, CRIM exhibited satisfactory performance in different treatment subgroups (Table S6).In the surgical resection group,CRIM achieved an AUC of 1.0 0 0 in the training cohort, and of 1.0 0 0 in the external validation cohort.In the TACE group, CRIM achieved AUCs of 0.986 and 0.960 in the training and external validation cohorts, respectively.In the ablation group, CRIM behaved with an AUC of 0.955 in the training cohort.Because patients chose liver resection as the main treatment in China, there were only three patients who fell into the ablation validation cohort.These three cases did not experience MaVI development.CRIM successfully recognized their non-MaVI status with 100% accuracy.

Independent risk factors and survival stratification

Peel9_fos_InterquartileRange was verified as a risk factor with a hazard ratio (HR) of 1.98 (P<0.001), which reflected the intensity interquartile range of the 9th inside layer of the tumor.With more severe wider interquartile range in the inside layer,the tumor behaved with higher intra-tumoral heterogeneity.The Cox-regression model could successfully divide patients into the high-risk and low-risk groups regarding the time of MaVI development (P<0.001).The median time of MaVI development was 5.0 months in the high-risk group and 18.5 months in the lowrisk group.The C-index for the Cox-regression model was 0.720[95% confidence interval (CI): 0.566-0.875]in the training cohort(n= 24), and 0.721 (95% CI: 0.569-0.871) in the external validation cohort (n= 24).Same prognosis stratifications were found on PFS (P<0.001) and OS (P= 0.002).The median PFS was 4.0 months in the high-risk group and 10.5 months in the low-risk group.The median OS was 12.0 months in the high-risk group and 30.5 months in the low-risk group.The Kaplan-Meier graphs re-garding the time of MaVI occurrence, PFS, and OS are shown in Fig.4.

Fig.4.The Kaplan-Meier curve manifested the risk factor Peel9_fos_InterquartileRange could successfully identify high-risk and low-risk subsequent MaVI occurrence patients ( P < 0.001, A ), meanwhile revealed its prognostic value on PFS ( P < 0.001, B) and OS ( P = 0.002, C).

Clinical usefulness of the developed model and open-public access

Decision curve analysis revealed that CRIM performed with improved net benefit of 20.0% in the training cohort with a cut-off possibility of 1%, and 22.0% in the external validation cohort at the cut-offpossibility of 2% (Fig.S5).

Discussion

In this multicenter study, we reported that 48 out of 226 patients developed MaVI after initial diagnosis, which would cause uncontrollable severe disease progression.To better monitor this situation and perform earlier intervention, CRIM was developed with the advantages of being noninvasive and low-cost.The proposed CRIM could discriminate MaVI with AUCs of 0.986 and 0.979 in the training and external validation cohorts, respectively.Peel9_fos_InterquartileRange was verified as an independent risk factor to predict the time of MaVI development.The cox-regression model achieved a C-index of 0.720 in the training cohort and 0.721 in the external validation cohort, which also indicated prognostic power to stratify patients regarding PFS (P<0.001) and OS(P= 0.002).These findings suggested that radiomics-based CRIM could assist in identification of MaVI development with prognostic implications in HCC management.

In clinical practice, the harmfulness of MaVI mainly depends on two aspects.Firstly, when HCC invades the portal vein and/or hepatic vein, it increases the risk of distant metastasis via the circulatory system, which is one of the HCC metastasis pathways [21–27].Secondly, MaVI occurrence would seriously jeopardize liver function, limiting most treatment applications.Relevant treatments require sufficient liver function, for which patients with MaVI are usually unqualified.Prevention and early identification of MaVI development are thus a necessary clinical task.

However, previous studies that analyzed preoperative predictors for vascular invasion paid more attention to MiVI [28–31].Although some vascular invasion studies included MaVI cases,they only focused on revealing the relevance between clinical, biological, or morphological variables and “existing”MaVI occurrence [ 32 , 33 ].Prediction of future MaVI development has not been explicitly explored.Mínguez et al.[34]reported a gene-expression signature that could distinguish existing vascular invasion with an accuracy of 0.69.Genetic research deepened the understanding of HCC, but such genetic-based tool requires tumor tissue when only partial patients receive resection treatment.For those who did not receive resection, additional biopsy would be required for genetic analysis, which may result in extra procedures, unnecessary patient suffering, as well as increased cost.In contrast, our proposed CRIM could provide MaVI prediction based on only routine radiological examinations.Besides, it surpassed the gene-expression signature with a high accuracy of 0.964 on the tough task to predict future MaVI development.

It is encouraging to see that our CRIM showed good generalization in this multicenter evaluation, reflecting its potential usefulness in a routine clinical setting in different hospitals regardless of CT scanning protocol and manufacturers.Furthermore, CRIM had maintained high-level performance in the stratification analysis of subpopulations regarding sex, BCLC staging,and AFP.

Beyond traditional radiomic features, we highlighted the usefulness of peel-offfeatures as defined based on the hypothesis of tumor growth.Peel-offlayers were defined as layers containing voxels from the outside to the inside of the tumor, shrinking in a 3-dimensional way.We hypothesized that MaVI occurrence might correlate with various degrees of invasiveness and growth patterns in these different peel-offlayers.As expected, the only radiomic feature that was selected to build both the radiomics signature and cox-regression model was exactly Peel9_fos_InterquartileRange.Peel9_fos_InterquartileRange reflected the intensity interquartile range of the inside layer of the tumor.With more severe intensity change in the inside layer, the tumor behaved with higher intra-tumoral heterogeneity.Pekar et al.[35]reported that intratumoral heterogeneity was positively correlated with tumor invasiveness but had a negative treatment response.Correspondingly,it inferred that HCC with larger Peel9_fos_InterquartileRange behaved with odious vascular invasiveness, difficult to control by the current treatment, accordingly, entailing a higher risk of subsequent MaVI occurrence.

The clinical factor node-in-node sign was incorporated into CRIM besides radiomic features.Theoretically, node-in-node sign was proved to be correlated with growth inconsistency within the tumor, which reflected the increased heterogeneity and also invasive tendency [36].The incorporation of clinical factors brought a slight AUC increase, demonstrating rationality to integrate both clinical/radiological and radiomic variables into the gathered network.

MaVI occurrence was not associated with treatment selection,which was revealed via the Pearson’s Chi-square test with aPvalue of 0.102.It inferred that MaVI development may likely relate to tumor substantial characteristics rather than the initial treatment manner.To further explore this issue, we conducted a treatment stratification analysis.The results were in accordance with the statistical results.This further confirmed that CRIM could recognize the essential difference between the MaVI group and non-MaVI group regardless of initial treatment manners.

The present study has limitations.Firstly, in this retrospective study, only CT imaging was involved in the analysis.Whether multi-modality radiomics such as CT plus MRI, especially MRI with gadoxetic acid, could provide more efficient information is worthy of exploration.Secondly, the analysis of MaVI occurrence time in this study was based on heuristic research.Although it manifested very promising potential on the time-predictable value of the radiomic features, considering that only 48 patients had subsequent MaVI occurrence time during the follow-up, further validation for MaVI occurrence prediction is needed in the next-step study.Thirdly, analysis of the location with high risks of MaVI, such as the portal vein vs.inferior vena cava, is under development.Finally, the manual segmentation to draw the tumor lesion was quite time-consuming.Semi-automatic segmentation methods and deeplearning methods are worth exploring in the future studies.

In conclusion, we have developed image-based models to allowa noninvasive and highly accurate assessment for MaVI development in HCC patients via a radiomics analysis.Our findings highlighted that the proposed CRIM model has potential robustness and inter-center generalization, which could better assist personalized treatment in HCC management.

Acknowledgments

None.

CRediT authorship contribution statement

Jing-Wei Wei:Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Software, Validation, Writing –original draft.Si-Rui Fu:Conceptualization, Data curation, Formal analysis, Writing –review & editing.Jie Zhang:Data curation, Formal analysis.Dong-Sheng Gu:Conceptualization, Formal analysis, Software, Writing –review &editing.Xiao-Qun Li:Data curation.Xu-Dong Chen:Data curation.Shuai-Tong Zhang:Formal analysis.Xiao-Fei He:Data curation.Jian-Feng Yan:Data curation.Li-Gong Lu:Funding acquisition, Supervision.Jie Tian:Conceptualization, Funding acquisition, Supervision, Writing –review & editing.

Funding

This work was supported by grants from the National Key R&D Program of China (2017YFA0205200, 2017YFC1308701,and 2017YFC1309100), National Natural Science Foundation of China (82001917, 81930053, 81227901, 81771924, 81501616,81571785, 81771957, and 61671449), and the Natural Science Foundation of Guangdong Province, China (2016A030311055 and 2016A030313770).

Ethical approval

The study was approved by the Ethics Committee of Zhuhai People’s Hospital (2018042601).Informed consent was waived for this retrospective study.

Competing interest

No benefits in any form have been received or will be received from a commercial party related directly or indirectly to the subject of this article.

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.hbpd.2021.09.011.