Preoperative Radiomics Analysis of Contrast-Enhanced CT for Microvascular Invasion and Prognosis Stratification in Hepatocellular Carcinoma

Introduction

The global burden of hepatocellular carcinoma (HCC) remains a grand challenge, with thirdly ranked mortality globally and even second-placed cause of cancer-related death in Asia-Pacific countries.^1,2 Surgical resection and liver transplantation (LT) are the potential radical treatments for HCC patients. But it is regrettable that five-year recurrence rate was approaching as high as 70% after surgical resection in HCC, and nearly 35% after LT, which exert unfavorable influences on their long-time prognosis.^3–5 Microvascular invasion (MVI) is regarded as one of the most crucial predictive risk factors for recurrence among HCC patients,^6,7 which is associated with complexity and heterogeneity of tumor. Some studies suggested MVI posed a great threat to poor outcomes of HCC patients.^8,9 MVI is defined as microscopic tumor invasion in smaller intrahepatic vessels, including the microvessels and lymphatics of the portal vein or hepatic artery.¹⁰ For macrovascular invasion, it can be identified through radiological examination such as ultrasonography (USG) and computed tomography (CT) before surgery, however, MVI is histopathologically diagnosed and confirmed through postoperative pathological examination. Hence, in case MVI status can be evaluated preoperatively, it would offer broad guidance to attain more widely clear surgical margins and may reduce postoperative recurrence rate.¹¹ Nevertheless, clinical diagnosis of MVI remains challenging during preoperative assessment. Although previous studies showed traditional imaging information, including tumor size, tumor multifocality and abnormal vascular perfusion were associated with MVI, these features were not widely adopted for diagnosis of MVI yet.^12,13 At present, no reliable methods can effectively diagnose MVI in the preoperative period apart from lesion biopsy.

The emerging and inspiring field of radiomics, which enables acquisition of comprehensive and quantitative features from medical images, may provide accurate diagnosis and prognosis information related to tumor complexity and heterogeneity.^14–16 Radiomics has showed the potential to be non-invasive imaging biomarkers for cancer diagnosis, differentiation of pathological subtype and assessment for treatment response or prognosis.^17,18 Despite its great potency, more reasonable and objective methods are needed for radiomics features quantification, and relationship between radiomics features and genomics factors needs for further exploration, so as to made precise clinical decision in the era of individualized oncology.¹⁴ Few studies have investigated prognosis of HCC patients with MVI based on radiomics signatures, when interpreting associations between biomarkers and radiomics signatures yet.

We aimed to investigate radiomics features based on contrast-enhanced CT (CECT) in HCC patients, and a novel nomogram incorporating radiomics signature and clinical information was constructed to predict MVI status as well as patients’ prognosis and validated in external data. Furthermore, quantitative real-time PCR (q-PCR) was applied to detect our previous research 10 tumor biomarkers from 53 randomly selected HCC patients. We assessed the relevance of gene relative transcript levels with radiomics signature of MVI, attempting to explore more underlying biological information of radiomics signatures.

Patients and Methods

Patients and Clinical Data

Our study was approved by the research ethics committee of Affiliated Hospital of Guilin Medical University and Peking University People’s Hospital, named Guilin dataset and Beijing dataset, respectively, and conformed to the principles of the Declaration of Helsinki. Informed consents were obtained from all subjects.

The diagnosis of HCC was based on clinical features, radiological examination, hepatic arteriography, alpha-fetoprotein (AFP) level, and postoperative pathological examination referring to the Clinical Diagnosis and Staging of Primary Liver Cancer established by the Professional Committee of Liver Cancer of Chinese Anti-Cancer Association. All enrolled patients underwent CECT scan within one month before surgery. Surgical resections were performed by well-trained and experienced hepatobiliary surgeons. Curative resection was defined as the complete removal of tumor, and no tumor cell invasions were seen in the resection margins by naked eye in more than twice follow-ups with no less than 4 weeks interval.¹⁹ To ensure complete removal of liver cancer, intraoperative abdominal USG and postoperative CT scan were routinely performed. All specimens were examined and diagnosed by two expert hepatopathologists ignorant about patients’ information, and a third pathologist was needed when results were discordant.

Follow Up

Patients were followed up at 1 month and 3 to 6 months after the operation. Routine postoperative examinations included physical examination, blood routine examination, liver function tests, coagulation function, serum AFP level, and abdominal USG. When any residual lesion or recurrence was suspected, CT or MRI scan would be performed immediately. The DFS time was calculated from date of surgery to date of recurrence, OS time was defined as date of surgery until date of death or last follow-up.

Radiomics Signature Acquisition

The acquisition of MVI-related radiomics signature involves several handcrafting steps: 1) Images acquisition and preprocessing; 2) Regions of interest (ROIs) selection; 3) Extraction of radiomics features from the ROIs; 4) Features dimension reduction and construction of radiomics signature (Rad-scores) (Figure 1A–G).

Figure 1 Workflow for the preoperative analysis of liver cancer based on CECT and clinical parameters. (A) CECT images acquisition and preprocessing. (B and C) We used the semiautomated editor to segment ROI, then 3D segmentation was reconstructed model. (D–F) Radiomics features about MVI status were extracted separately from ROI and visualized, LASSO was used to select the optimal characteristic features by 10-fold cross-validation, ROC and AUC were used to evaluate the feasibility for the detection of MVI status by above-selected features. (G) Calculate Radiomics signature. (H) Construct predicting models based on Radiomics signature and clinical parameters. (I) Performance evaluation of predicting models. (J) Create the nomogram based on outstanding performance. Abbreviations: CECT, contrast-enhanced computed tomography; 3D, the three dimensional; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; AUC, area under the curves; ROI region of interest; MVI, microvascular invasion

Image Acquisition and Preprocessing

Each patient’s CECT images were acquired according to the AASLD guidelines.²⁰ CECT scans were conducted by multidetector CT scanners: GE Light Speed 64-, 128-, and 256 slice spirals. Arterial phase (AP) imaging was performed by using contrast medium tracking. The trigger is automatically located in lower border of thoracic aorta, threshold: 180 HU. A delay of 5 seconds before image acquisition. The venous phase (VP) started on 30s delay after the end of AP scanning. The CECT scanning parameters were 120 kVp, automatic tube current modulation (mA), Noise Index: 8, and 1.25 mm interval. Images in AP and VP were standardized using z-score normalization to obtain standard normal distribution of image intensities and resampled to voxels of 1 x 1×1 m³ through the Python (v3.7, https://www.python.org) and the open source Simple-ITK package.

ROIs Selection

The CECT images of each patient were conducted using an open-source software, 3D Slicer (v4.11, https://www.slicer.org). ROIs were segmented from tumor in AP and VP phases using the semiautomated editor module. When patients had multiple tumors, the largest one was analyzed. Two radiologists with 8- and 10-years’ experience reviewed the segmented images, respectively (reader1 and reader2). If results were discordant, a third pathologist with 10-years’ working experience joined to confirm the results, and the final ROIs were determined by majority of vote. Readers were blinded to clinical and pathological parameters.

The Extraction of Radiomics Features

The various radiomics features were extracted separately from each ROI on AP and VP via an open-source package named PyRadiomicswith different background filters, radiomics features from CECT were selected.²¹

Features Dimension Reduction and Radiomics Signature Construction

Value of radiomics features was standardized using Z-score in the training cohort (z = (x − μ)/σ). The optimal radiomics parameters were selected by the least absolute shrinkage and selection operator (LASSO) with 10-fold cross-validation in the training cohort and were applied to validation cohort. Radiomics score (Rad-score) was calculated for each patients via a linear combination of the selected features by their respective coefficients, which we defined as radiomics signature. The receiver-operating characteristic (ROC) curve and area under the curves (AUC) were used to evaluate the discrimination of radiomics signature as a diagnostic variable for detection of MVI status.

Constructing Prediction Models and Performance Evaluation

Three models for preoperative prediction of MVI status were established: radiomics model (RM), clinical model (CM) and clinical-radiomics model (CRM). CM excluded the radiomics signatures, while RM was built on just radiomics signature. CRM was constructed based on logistic regression analysis according to Akaike Information Criteria (AIC) and statistical significance.

Univariate logistic regression analysis was performed to assess risk factors for discriminating MVI status. Variables with p-value <0.05 were incorporated into the multivariate logistic regression analysis. AUCs and the Harrell’s concordance index (c-index) were used to compare discrimination ability of the three models in both training cohort and validation cohort. Calibration curves were plotted, and decision curve analysis was considered to determine potential clinical decision thresholds. A nomogram based on the CRM was built.

Survival Risk Stratification of HCC Patients

The risk score of each patient was calculated by nomogram. Based on median score of each cohort, patients were divided into the high and low risk subgroups. Kaplan–Meier survival curves were used to estimate overall survival (OS) and Disease-free survival (DFS), and Log rank tests were used to compare survival time between the subgroups.

Biomarker Assessment of HCC Tissues by Quantitative Real-Time PCR (q-PCR)

53 individuals were randomly selected among all patients, and their tumor tissues were performed by quantitative real-time PCR (q-PCR) analysis to assess gene relative transcript levels. 10 biomarkers were selected, including BARD1, CKAP5, MAP17, AKR1B10, MDR1, ABCC3, AR, BCAT1, DTGCU2, DOCK3 and serum AFP. Experimental methods, primer sequences of these biomarkers and the reference were showed in Supplementary Table S1. To determine the association between our radiomics signature and these biomarkers with serum AFP level, we unified and normalized all data and then calculated Pearson's correlation coefficient (PCC) and associated p-values, absolute value of PCC greater than 0.3 were considered statistically significant (p < 0.5).

Statistical Analysis

All statistical analyses were performed using software R (version 4.11, https://www.r-project.org/). Kolmogorov-Smirnov test checked whether the continuous variables were normally distributed. Normally distributed and non-normally distributed variables were expressed as mean ± standard deviation or as median with interquartile range (IQR), respectively. Continuous variables were analyzed using the Student’s t-test or Mann-Whitney U-test, categorical variables were tested using Pearson χ² test or Fisher exact probability test. The threshold for statistical significance was set as p < 0.05.

Results

Baseline Clinical Characteristics of HCC Patients

Guilin dataset was retrospective collected between June 2012 to December 2017, Beijing dataset was retrospective collected between August 2013 to June 2017. HCC patients with or without MVI were selected for subsequent analyses.

A total of 421 patients (male, n = 357; female, n = 63) were included in the study and MVI-present accounted for 43% cases (n = 182). Then 221 patients were excluded from the study according to inclusion and exclusion criteria, and 295 patients from the Guilin dataset were divided into the training cohort, and 126 patients from the Beijing dataset were divided into validation cohort. There were no statistically significant differences of clinical and pathological characteristics between the two cohorts (Table 1). Patient enrollment process and inclusion-exclusion criteria were shown in Figure 2.

Table 1 Comparison of HCC Patients Between Training and Validation Groups

Figure 2 Data diagnosed hepatocellular carcinoma with/without MVI collected process.

Radiomics Features and Foundation of Radiomics Signature

ROIs were delineated from AP and VP and a total of 1962 radiomics features (981 features for AP image and 981 features for VP image) were extracted. In training cohort, LASSO regression selected 11 MVI-related radiomics features, with 6 from AP and 5 from VP (Figure 1E). Radiomics score was calculated for each HCC patients based on identified features and their respective coefficients. The radiomics features demonstrated excellent discriminating ability, with an AUC of 0.796 and 0.810 in the training and validation cohort, respectively (Figure 1F). The formula of radiomics signature and results of selected features were shown in Supplementary Table S2.

Univariate and Multivariate Logistic Regression Analyses

We performed univariate logistic regression analysis using the clinical parameters and radiomics signature in training cohort and results were shown as follows: Fibrinogen (odds ratio [OR], 2.02; 95% CI, 1.26–3.22; p < 0.001), Prealbumin (OR, 0.51; 95% CI, 0.31–0.84; p = 0.015), AST (OR, 3.16; 95% CI, 1.96–5.11; p < 0.001), ALT (OR, 1.98; 95% CI, 1.23–3.17; p < 0.001), GGT (OR, 2.79; 95% CI, 1.60–4.87; p < 0.001), AFP levels (OR, 1.70; 95% CI, 1.07–2.71; p = 0.022), tumor size (OR, 4.38; 95% CI, 2.61–7.36; p < 0.001), tumor number (OR, 2.80; 95% CI, 1.60–4.88; p < 0.001), radiomics signature (OR, 1.08; 95% CI, 1.05–1.10; p < 0.001). The above significant indicators were incorporated into the multivariate logistic regression analysis, and AST (OR, 2.34; 95% CI, 1.376–3.982; p < 0.001), tumor number (OR, 2.86; 95% CI, 1.563–5.325; p < 0.001), radiomics signature (OR, 1.07; 95% CI, 1.044–1.097; p < 0.001) were identified as significant independent predictors for HCC patients with MVI (Table 2).

Table 2 Univariate and Multivariate Logistic Regression Analysis of HCC Patients with Microvascular Invasion in Training Group

Development and Validation of the Nomogram

The multivariate logistic regression analysis selected AST, tumor number, radiomics signature as significant factors for MVI prediction. Through AFP level (OR, 1.270; 95% CI, 0.747–2.156; p = 0.376) was not included, we still incorporated it in prediction model as it remained well acknowledged and significant for HCC patients clinically. Accordingly, we defined AST, tumor number and serum AFP level as clinically MVI relevant parameters, while radiomics signature as radiological one. Then radiomics model (RM), clinical model (CM) and clinical-radiomics model (CRM) were established for preoperative prediction of MVI status. The AUCs of CRM, RM and CM were 0.767 (95% CI: 0.713–0.822), 0.724 (95% CI: 0.666–0.782) and 0.697 (95% CI: 0.637–0.757) in the training cohort and were 0.793 (95% CI: 0.714–0.874), 0.750 (95% CI: 0.666–0.834) and 0.648 (95% CI: 0.552–0.744) in the validation cohort, respectively (Figure 3A and B).

Figure 3 ROC analyses to assess the abilities of the three models to predict MVI in the (A) training and (B) validation cohorts. Calibration curves of the three models in the (C) training and (D) validation cohorts. Decision curve analysis of the nomogram in the (E) training and (F) validation cohorts. Solid line: predictive nomogram. The predicted probabilities of MVI are plotted on the X-axis, and actual MVI probabilities are plotted on the Y-axis. Abbreviations: MVI, microvascular invasion; ROC, receiver operating characteristics.

Calibration curves were used to further assess the accuracy of the three models, and the calibration curve of CRM indicated a good predictive performance in both training and validation cohort (Figures 3C and 4D). Decision curve analysis indicated CRM added more net benefit than the other two models, especially in probability threshold of 40–68% and 46–76% in both two cohorts, which suggested that CRM had significant clinical utility (Figure 3E and F). Ultimately, we constructed a novel nomogram based on CRM (Figure 1J). The nomogram is ideal for predicting MVI status among HCC patients, which has favorable performance and high precision with C-index of 0.767 (95% CI: 0.713–0.822) and 0.793 (95% CI: 0.714–0.874) in training and validation cohorts, respectively.

Figure 4 Survival curves according to nomogram. The Kaplan–Meier curves depict disease-free survival in HCC patients according to the classification of nomogram in the (A) training and (B) validation cohorts.

Prognostic Stratification for Survival

We calculated the total score for each patient based on nomogram in the training and validation cohorts, respectively. Using the median risk score as cut-off threshold, patients were divided into the low- and high-risk subgroups: for the training cohort, the median OS time was 71 months, 33 months and the median DFS time was 60 months, 18 months in the low- (score ≤35.758) and high-risk (score >35.758) subgroups, respectively; and in the validation cohort, the median OS time was 70 months, 25 months and the median DFS time was 36 months, 13 months in the low- (score ≤38.303) and high-risk (score >38.303) subgroups, respectively.

Kaplan–Meier analysis showed significantly different DFS and OS curves of the low- and high-risk subgroups in both training cohort (log rank test, p < 0.0001) and validation cohort (log rank test, p < 0.0001), respectively (Figure 4A and B, Supplementary Figure S1A and S1B).

Correlation Analysis Between Biomarkers and Radiomics Signature

53 patients were randomly selected in our study, and 10 biomarkers were detected by q-PCR analysis from the tumor tissues. Biomarkers’ relative transcript levels, serum AFP level and radiomics signature were combined for the correlation analysis (Figure 5), and results showed that BCAT1 (PCC = 0.335, p =1.542e-2), DTGCU2 (PCC = 0.362, p = 8.415e-3), DOCK3 (PCC = 0.352, p =1.0511e-2) and AFP (PCC = 0.382, p = 5.225e-4) had significant association with radiomics signature.

Figure 5 Association of radiomics signature with biomarker expression levels. Circle plot visualizations of correlation between radiomics signature and biomarkers expression levels are shown in (A). Scatter plot visualizations of correlation between radiomics signature and biomarker expression levels and serum AFP levels are shown in (B–E).

Discussion

Consistently, the presence of MVI is considered as a significant risk indicator for worse outcomes in HCC.^6–9 The purpose of our study is to develop and validate a individualized nomogram for estimating MVI status preoperatively based on combination of radiomics signature with routine clinical variables, including AST, tumor number and AFP level. In addition, the new nomogram enables favorable stratification of HCC patient into different risk groups. Final, we explored correlations between radiomics signature of MVI based on CECT and 10 tumor biomarkers from HCC issue measured by q-PCR with serum AFP level.

As the first-line treatment for HCC, operation procedure of surgical resection is determined by surgeons at present. Different strategies would be applied based on patients’ conditions in clinic, non-anatomic resection would be conducted if tumor were poorly differentiated or with MVI, while anatomic resection was preferred to non-anatomic resection and may associate with lower rate of recurrence.²² Therefore, preoperative assessment of probability of MVI is of tremendous significance and may help improve prognosis of HCC patients dramatically. Nevertheless, preoperative MVI assessment remained a conundrum. In our study, the CRM in predicting MVI exhibited outstanding performance, better than the other two models, with AUCs of 0.796 and 0.810 in two cohorts. The calibration curve of CRM also demonstrated good consistency among three models, suggesting CRM may be a reliable tool for estimation of MVI. Moreover, the decision curve analysis showed that our nomogram had the highest clinical net benefit across most ranges of threshold probabilities, which further demonstrated good clinical application value and discriminative performance of CRM.

It is still noteworthy that preoperative AFP level showed deficient ability to predict MVI status in the multivariate regression analysis. In order to matching general strategies, AFP levels will be excluded from the final regression model. However, the effect of a predictor may depend on the other predictors or contain nuances unique to the full dataset.²³ The association was not statistically significant may not necessarily mean it was not significant in the chain of causation. AFP, performing as a well-acknowledged biological predictor of HCC, reveals burden of tumor and aggressive phenotype.^24,25 Hence, AFP was still considered as independent predictor of MVI status in the modeling procedures. Tumor number is often regarded as a surrogate marker for MVI. It has been shown previously that tumor number positively related to the incidence of MVI, and frequency of MVI dramatically increased as tumor number increased.^26–28 Elevated AST level has been investigated as a detrimental factor on prognosis for HCC patients, which reflects liver burden and may indicate anti-aggressiveness of liver to some extent.^29,30 To our knowledge, few studies found AST as an independent predictor of MVI, AST was identified as independent predictor of MVI for HCC patients in our current study.

Furthermore, among all independent risk predictors in CRM, the weight of radiomics signature constituted the largest proportion, suggesting that radiomics features may contain crucial information about MVI and great potential to effectively predict survival and prognosis of HCC patients, which emphasized radiomics features’ great potential in clinical application. Kim et al found that radiomics features extracted from Gadoxetic Acid-Enhanced Magnetic Resonance Imaging could be used to predict early recurrence of patients with HCC after surgical resection.³¹ Xu et al based on interested tumor regions, derived three different ROIs through the erosion and dilation algorithm and used these ROIs in their study. The AUCs of the models in predicting MVI status were as high as 0.909 and 0.889 in the training and validation cohort, respectively.³² However, their models included much more vital factors than Rad-scores: tumor capsule, tumor margin, peritumoral tissue. And above-mentioned analyses did not reveal biological significance behind quantitative radiological characteristics.

Different from semantic features, such as tumor number, directly observed in the preoperative images by radiological specialists, our radiomics signature is obtained from CECT images via specific computer-aided methods other than naked eyes. Radiological features could reflect subtle texture changes in neoplastic lesions that were hardly discernible to human eyes, some of which may also contain significant implication about microenvironment of tumor and its periphery.

We also notice that before radiomics signature production in our study, the predictive performance of the result of features dimensionality reduction better than the nomogram, this suggesting the existence of issues that a poor reproducibility and comparability among the high-throughput features. Rad-scores can be drawn from a diversity of algorithms, there may be potential limitations of LASSO regression for burgeoning high-throughput features. Currently, authoritative and universal algorithms are needed to settle the mentioned problems. To improve feature reproducibility, one important method is to calculate the objective Rad-scores, which can help explain significance of radiomics signature and predict biological characterization.

However, the physiological mechanism behind the radiomics signature is not fully understood at present. We hypothesize that radiomics signature of MVI in our study is associated with expression of important biomarkers in the pathogenesis of HCC. Further researches are necessary to confirm the hypothesis. Then, we carried out q-PCR analysis for 10 important tumor biomarkers of HCC, whose roles may be involved in metabolism relevant events of cells and inhibition of tumor cell proliferation. We found negative and positive associations between the expressions of those biomarkers and radiomics signature. Correlations showed their association were real existence. BCAT1 may be a significant biomarker of tumor metastasis in HCC and its overexpression may induce circulating tumor cells release via triggering epithelial–mesenchymal transition.^33,34 In HCC, DTGCU2 is elevated, and its high expression is significantly correlated with worse prognosis and metastases.³⁵ DOCK3 acts as a cancer‐promoting gene, contributing to adhesion, migration and invasion of tumor cells.^36,37

Previous findings revealed that MRI radiomic features may play important roles to distinguish Glioblastoma subtypes, become potential biomarkers of cancer, and even involve in some important molecular pathways.¹⁷ Some scholars demonstrated radiomics features extracted from CT exhibited substantial heterogeneity and could predict survival for lung cancer, head and neck cancer patients, and were associated with gene-expression profiles or expression patterns.¹⁸ It may be significant differences in image features among differences populations, which contained valuable information for prognosis prediction, and even some image features were linked with gene expressions at the protein level.³⁸ Whereas only unimodal analytic approaches were used in the above-mentioned studies. Due to the difficulty of gathering multi-modal data, it is very hard to investigate image features comprehensively. These preliminary experiments illustrated associations between tumor biochemistry and medical image features, and that quantitative image features were of sufficient potential in predicting tumorigenesis, development and prognosis of many cancers and even possible to improve medical decision support. Nevertheless, relationships between radiomics features and clinical factors, gene expression levels and other biological behaviors are still needed to be further explored.

Several limitations of our present study should be acknowledged. To begin with, this was a retrospective analysis along with well-known and inevitable limitation of retrospective data, though the data were screened carefully. Secondly, the CECT images were collected from different time and institution, albeit complying with a standard research protocol, we cannot deny the fact that the results may ascribe to diverse institutions and scanners. What’s more, with the progression of multimodality imaging techniques, 18F-FDG PET/CT and gadoxetic acid disodium-enhanced MRI play an important role in clinical HCC’s examination.^39,40 Here we merely explore one type of imaging data. Fourthly, the number of genes and tissue samples tested by qPCR is limited. Finally, we have targeted only one kind of ROIs from tumor region, which may omit potential in the peritumoral area. Based on the entire tumor profile, utilizing complex algorithms, automated rebuilding of a new ROI, which is decreased to 50% of the initial volume, the classification models may perform well to predict MVI status.³² Nonetheless, these studies indicated that significant information may be contained in different tumor regions of HCC.

Conclusion

We demonstrated a novel nomogram based on AST, tumor number, AFP level and radiomics signature to predict MVI and prognostic risk of HCC. Our nomogram is easy to calculate and contains variables routinely assessed in clinical examinations. This study provides some biologic corroboration to radiomics signatures of MVI at the same time. Meanwhile we expect to optimize and standardize an imaging model to broaden the application of radiomics for cancers.