Home
JournalsCollections
For Authors For Reviewers For Editorial Board Members
Article Processing Charges Open Access
Ethics Advertising Policy
Editorial Policy Resource Center
Company Information Contact Us
OPEN ACCESS

Associating Preoperative MRI Features and Gene Expression Signatures of Early-stage Hepatocellular Carcinoma Patients using Machine Learning

  • Xiaoming Li1,#,
  • Lin Cheng1,#,
  • Chuanming Li1,
  • Xianling Hu1,
  • Xiaofei Hu1,
  • Liang Tan2,3,
  • Qing Li4,
  • Chen Liu1,*  and
  • Jian Wang1,* 
Journal of Clinical and Translational Hepatology   2022;10(1):63-71

doi: 10.14218/JCTH.2021.00023

Received:

Revised:

Accepted:

Published online:

 Author information

Citation: Li X, Cheng L, Li C, Hu X, Hu X, Tan L, et al. Associating Preoperative MRI Features and Gene Expression Signatures of Early-stage Hepatocellular Carcinoma Patients using Machine Learning. J Clin Transl Hepatol. 2022;10(1):63-71. doi: 10.14218/JCTH.2021.00023.

Abstract

Background and Aims

The relationship between quantitative magnetic resonance imaging (MRI) imaging features and gene-expression signatures associated with the recurrence of hepatocellular carcinoma (HCC) is not well studied.

Methods

In this study, we generated multivariable regression models to explore the correlation between the preoperative MRI features and Golgi membrane protein 1 (GOLM1), SET domain containing 7 (SETD7), and Rho family GTPase 1 (RND1) gene expression levels in a cohort study including 92 early-stage HCC patients. A total of 307 imaging features of tumor texture and shape were computed from T2-weighted MRI. The key MRI features were identified by performing a multi-step feature selection procedure including the correlation analysis and the application of RELIEFF algorithm. Afterward, regression models were generated using kernel-based support vector machines with 5-fold cross-validation.

Results

The features computed from higher specificity MRI better described GOLM1 and RND1 gene-expression levels, while imaging features computed from lower specificity MRI data were more descriptive for the SETD7 gene. The GOLM1 regression model generated with three features demonstrated a moderate positive correlation (p<0.001), and the RND1 model developed with five variables was positively associated (p<0.001) with gene expression levels. Moreover, RND1 regression model integrating four features was moderately correlated with expressed RND1 levels (p<0.001).

Conclusions

The results demonstrated that MRI radiomics features could help quantify GOLM1, SETD7, and RND1 expression levels noninvasively and predict the recurrence risk for early-stage HCC patients.

Keywords

Hepatocellular carcinoma, Gene-expression, MRI, Radiomics feature

Introduction

HCC is among the leading causes of cancer-related death, with an estimated 42,030 new cases and 31,780 mortalities each year in the USA.1 Although the overall survival (OS) has improved up to 18% in the last four decades, the mortality rate has doubled, and the incidence rate has tripled during the same period.2 Hepatic resection and locoregional ablation are curative treatments that improve the OS of HCC patients. However, HCC recurrence, a key contributor to reducing the long-term survival rate, is observed up to 60–70% in the patients after surgical resection.3,4

Genomic profiling of tumors has revealed a complex genetic landscape and characteristics of HCC; moreover, it has identified key genes that play roles in tumorigenesis and tumor growth.5–7 The identified genes can be used to discriminate heterogeneous responses to treatments and to predict HCC prognosis.8–10 Several genes are already identified as associated with the prognosis of HCC patients, as reported from previous studies in the literature.11–13 Komatsu et al.11 linked poor prognosis of HCC patients with reduced Rho family GTPase 1 (RND1) gene expression, which was associated with increased HCC cell proliferation, invasion, and chemoresistance to cisplatin. Golgi membrane protein 1 (GOLM1) was associated with tumor growth and metastasis, while being implicated as a cause of early recurrence and poor survival.12 Moreover, a previous study revealed the affiliation of decreased level of the SET domain containing 7 (SETD7) protein, which regulates cell cycle, with HCC recurrence and metastasis, and therefore poor prognosis.13

Radiomics has great potential to describe the structural characteristics of the tissue by quantitative analysis of non-invasive medical image data using different mathematical perspectives. In the previous studies, radiomics features were utilized to predict disease stages,14 clinical outcomes,15 treatment responses,16 tumor heterogeneities,17 and early detection of cancers.18 Recently, several studies utilized radiomics features to investigate the relationship between gene expression signatures and medical imaging data for several cancers, including glioblastoma multiforme,19 breast cancer,20 non-small cell lung cancer,21 and HCC.22–24 Although significant correlations have been observed between imaging data and gene expression signatures, the correlation between quantitative magnetic resonance imaging (MRI) features and genes associated with disease recurrence in early-stage HCC patients still remains the primary focus in cancer research.

In this study, we developed multivariable regression models by interpreting underlying structural characteristics from T2-weighted MRI of tumor tissues and investigated whether radiomics analysis of preoperative MRI data can assist noninvasive quantification of gene expression signatures associated with disease recurrence in early-stage HCC patients.

Methods

Patients

Written consent was obtained from all study participants using protocols approved by the Institutional Research Subpanel on Human Studies at Southwest Hospital. The patients were selected according to pre-determined inclusion criteria: I) having Barcelona Clinic liver cancer stage 0 or A; II) having pathologically-verified HCC; III) having undergone preoperative MRI; IV) having undergone tumor surgical resection; and V) having gene expression signatures for resected tumor tissue. A total of 92 patients that visited our institution between October 2014 and September 2016 were included in this retrospective study.

Genomic profiling

The microarrays were generated from formalin-fixed, paraffin-embedded tissue specimens collected from surgically-resected tumors, and immunochemistry analysis was performed according to the commercial kits’ manufacturers’ protocols (GOLM1: A12584, ABclonal Technology, Woburn, MA, USA; SETD7: ab14820, Abcam, Cambridge, UK; RND1: ab206669, Abcam). The gene expression scores were evaluated by an expert pathologist and demographic information was collected from unidentified patient data. Gene expression signatures and other information of the patients are presented in Tables 13.

Table 1

Demographic information of the patients regarding GOLM1 gene expression levels

GOLM1 gene expression levels
123456
Age in years6849±13.550.46±11.5253±10.440±1275
Gender
Male214183421
Female016610
Liver cirrhosis
Positive05132321
Negative210111710
HBsAg
Positive112203421
Negative134600
Blood test
AFP2.65±1.32197.46±318.81160.63±262.56121.8±228.2266.7±461.9800±0
ALP90±1.2497.2±26.6092.08±41.4388.8±30.989.7±31.8101±0
GGT38.5±19.0940.87±25.0450.83±45.0867.7±56.242.3±3232±0
AST36.6±4.81144.09±159.05123.65±133.34144.9±189.6141.3±68.620±0
ALT40.6±2.26150.23±161.73115.80±141.68157.2±226.7148.5±127.913±0
AFU25.85±4.7431.88±10.1030.88±11.3731±11.125.1±10834.4±0
5′-NT5.35±1.917.14±4.206.69±4.057.7±4.16.1±2.17.7±0
ADA9.1±1.7013.87±6.4313.69±6.5212.9±6.217.7±615.4±0
Table 2

Demographic information of the patients regarding SETD7 gene expression levels

SETD7 gene expression levels
024
Age in years51.8±1258.4±10.948±0
Gender
Male6660
Female1112
Liver cirrhosis
Positive4221
Negative3551
HBsAg
Positive6451
Negative1221
Blood test
AFP171.1±277.16.7±5.8402.3±562.5
ALP90±32.2102.7±40100±19.8
GGT56.8±49.250.1±37.329.5±14.9
AST121.4±142.6232±313.640.2±11.1
ALT132.5±167.1228.8±350.322.9±10
AFU30.8±1129±2.227.6±18
5′-NT7.2±4.17±35±0.8
ADA13.4±6.213.3±7.114.1±7.7
Table 3

Demographic information of the patients regarding RND1 gene expression levels

RND1 gene expression levels
012345
Age in years50.9±10.653.9±11.754.7±11.447±1.437.5±13.463
Gender
Male191819231
Female813010
Liver cirrhosis
Positive111115111
Negative1687130
HBsAg
Positive221516231
Negative546000
Blood test
AFP237.9±328.594.4±195.1127.3±238.151.9±683.74±3.59.9
ALP90.2±32.194.6±3488.3±30.162.5±19.175.9±44.9110
GGT41.5±25.885.7±80.951.7±35.650.5±48.854±22.664
AST87.9±129164.8±218.1104.9±106.7407.7±185.8106.7±7337
ALT90.2±141.5176.8±231.9105.4±109.5323.4±78.6101.9±86.141
AFU28.3±7.935±13.828.3±7.528.6±3.737.8±17.333
5′-NT5.7±2.38.7±5.36.7±4.64.7±37.4±1.48.1
ADA12.9±5.915.1±5.813.3±4.98.5±8.913.9±8.313.7

MRI image acquisitions

Patients were imaged in our institution using a 3T MRI scanner (MAGNETOM Trio; Siemens Healthcare, Erlangen, Germany) with a 18-ch body coil. Preoperative T2-weighted MRI data were acquired with a two-dimensional half Fourier acquisition single shot turbo spin echo sequence (“HASTE”) under breath-holding with imaging parameters: TR/TE: 1,000/88 ms, slice thickness: 6 mm; section gap: 7.8 mm; matrix size: 244×320. Two radiologists (Observer 1: C Liu, with 10 years of experience in abdominal imaging; Observer 2: X Li, with 6 years of experience in abdominal imaging) were blinded to other clinical information, and they independently selected the slice with the biggest tumor diameter on T2-weighted images. Interobserver variation was measured with statistics. The tumor diameter was measured 3 times and recorded as mean value by Observer 1.

Feature extraction and selection

Afterward, tumor tissues were outlined on MRI slice to generate a region of interest (ROI) using ITK-SNAP (v.3.6.0), by the two observers. Observer 1 repeated tumor segmentation twice in a week and Observer 2 independently performed the segmentation to evaluate test-retest and interobserver reproducibility. The reproducibility was subject to the intraclass correlation coefficient. An example of T2-weighted MRI is presented in Figure 1 (panel A) with an enlarged HCC tumor shown in panel B. After tumors were outlined, the MRI image was resampled to 1×1 mm using a B-spline interpolation approach to match in-plane image resolution for consistency during tumor characterization with quantitative texture analysis.25 Afterward, MRI signal intensity was standardized by performing a z-score normalization method. The standardized image intensities were discretized using a fixed-bin size approach as preferred by Image Biomarker Standardization Initiative.26 Four different bin sizes (8, 16, 32, and 64), as the level of image specificity, were evaluated to find the discretization parameters that can perform the best interpretation of the association between tumor heterogeneity and gene expression levels. A representative MRI slice quantized with different levels (image specificity) is presented in Figure 1C–F.

A representative MRI slice for an early-stage HCC patient.
Fig. 1  A representative MRI slice for an early-stage HCC patient.

(A) Original MRI slice (arrow). (B) Enlarged tumor. (C–D) Effect of image quantization performed before feature extraction. The same MRI slice is quantized with 8 (C), 16 (D), 32 (E), and 64 (F) levels (arrow). HCC, hepatocellular carcinoma; MRI, magnetic resonance imaging.

A total of 307 radiomic features regarding intensity, texture, pattern, and shape characteristics of the images was computed from ROIs of the selected slice in T2-weighted MRI using 10 different feature extraction approaches and two transformations (wavelet and gradients) using MATLAB (v.9.1; MathWorks, Natick, MA, USA). The feature extraction methods were: first-order statistics (FOS), with six features; gray-level co-occurrence matrix (GLCM), with nine features; gray-level run-length matrix (GLRLM), with thirteen features; gray-level size-zone matrix (GLSZM), with thirteen features; neighborhood gray-tone difference matrix (NGTDM), with five features; local binary patterns (LBP), with ten features; fractal analysis, with one feature; (8) shape metrics, with nine features; and moments, with forty-three features.16,27–29 In addition, six FOS features and one hundred ninety-two features (FOS, GLCM, GLRLM, GLSZM, variance, and power) were computed from oriented gradients histogram and four sub-band wavelet images, respectively. GLCM and GLRLM features were computed through four main directions (0°, 45°, 90°, and 135°) and combined by averaging. Afterward, all the features were normalized to the range of 0 and 1 by utilizing min-max normalization. The complete list of features is presented in Supplementary Table 1.

The trivial features were identified by performing a multi-step feature selection process. Initially, features were selected by their correlation with tumor size (r>0.75). Afterward, features were clustered based on their cross-similarity (r>0.75) measured with Pearson correlation coefficient to minimize artificially increased numbers of candidate features. In each cluster, correlations of each feature pair were computed and the one with the largest average similarity was selected as the representative of the cluster. Later, the interactions of representative features were assessed using the RELIEFF algorithm with 10 nearest neighbors to determine feature importance/rank.30 The top-ranked 20 features were selected as the final candidate-feature set to associate with gene expression signatures. The feature selection procedure was repeated separately for each level of intensity (8, 16, 32, and 64) and key variables were determined separately for each gene.

Statistical analysis

The identified features with the RELIEFF algorithm were further examined in the exhaustive search experiment, while the features were applied to generate regression models using a support vector machines approach with a radial basis function kernel and 5-fold cross-validation. Spearman’s rank correlation coefficient was used to evaluate the performance of the generated regression models. Wilcoxon rank-sum test was utilized to analyze continuous variables. Meanwhile the chi-squared test was performed to analyze dichotomous variables. Moreover, correlation analysis was performed to determine the association between patient clinical characteristics and gene expression signatures. The continuous variables were presented as mean±standard deviation (µ±σ) and p<0.05 was approved as statistically significant.

Results

Patients

A total of 92 patients were incorporated in our retrospective study, which includes 14 female patients (51.36±13.45 years) and 78 male patients (52.17±11.67 years) with early-stage HCC. The average size of the lesions was 3.9±1.5cm (range: 1.1–6.3cm). The correlation analysis demonstrated that patient characteristics including blood markers were not strongly correlated with any of the gene expression signatures (|r|<0.254). For the GOLM1 gene, cirrhosis had the highest correlation value (r=0.234, p=0.027) and gender had the strongest association with the SETD7 gene signature (r=−0.254, p=0.018). Moreover, AFP was negatively correlated with the expression level of the RND1 gene (r=−0.232, p=0.052). The agreement between the two observers was excellent (=0.78) for the chosen slice with the biggest diameter of tumor tissue.

Feature selection

The key features were identified in a multistep feature selection procedure. First, features with high stability (intraclass correlation coefficient >0.75) from both observers were kept for further analysis. An average of 33 variables (35 to 31 features), strongly associated with tumor size (r≥0.75), was removed from the feature set. With decreasing number of features correlated with tumor size, the image specificity was improved, except for a bin size of 64 in which 32 features were included. Later, remaining features were grouped into an average of 41 clusters (44 to 35 groups) based on their pairwise similarity measured by Pearson correlation coefficients. However, we did not observe the same behavior with features computed from MRI data quantized with 64 levels (40 groups). The correlations between representative features of each cluster and gene expression signatures were visualized in Figure 2, which shows the need for multivariate models. The correlation between features and gene expression were seen in Supplementary Table 5.

Correlation of gene expression signatures of GOLM1, SETD7, and RND1 and textural features extracted from MRI data with different quantization levels.
Fig. 2  Correlation of gene expression signatures of GOLM1, SETD7, and RND1 and textural features extracted from MRI data with different quantization levels.

The statistically significant features are highlighted in green and features with a p-value between 0.05 and 0.10 are shown in gray. Multivariate analysis is required to improve the association between gene expression signatures and MRI image features. GOLM1, Golgi membrane protein 1; MRI, magnetic resonance imaging; RND1, Rho family GTPase 1; SETD7, SET domain containing 7.

The inter-relationship of representative features is demonstrated in Figure 3A. The representative features were ranked by the RELIEFF algorithm via analysis of gene expression signatures. The top-ranked 20 features were selected for multivariable regression models. The candidate variables computed from MRI with the lowest image specificity for GOLM1, SETD7, and RND1 genes are presented in Figure 3 (panels B–D). Other candidate features are presented in Supplementary Tables 24.

Evaluation of the features to determine the candidate feature list for multivariate regression models.
Fig. 3  Evaluation of the features to determine the candidate feature list for multivariate regression models.

(A) Pairwise relationship of the representative features computed from 8-level MRI data. (B–D) The top-ranked 20 features of 8-level MRI data that was determined for GOLM1 (B), SETD7 (C), and RND1 (D) genes. GOLM1, Golgi membrane protein 1; MRI, magnetic resonance imaging; RND1, Rho family GTPase 1; SETD7, SET domain containing 7.

Model evaluation

The image specificity to predict gene expression levels were evaluated by measuring the association between the selected features and gene expression signatures using multivariable regression models. In Figure 4, the relationship between the overall behavior of the models and the number of features was presented separately for three genes. For the training data, the correlation between multivariate regression models’ response and gene expression scores was improved with an increasing number of features. However, the correlation did not improve similar level for the validation data. r values for training and validation differed with increasing model complexity. Thus, the final regression models were generated with features determined by the overall performance of the circular training and validation experiments. Besides, there was no statistically significant difference between performances of the regression models with different image specificities (p>0.3). The final multivariate regression models were generated with key features and optimal level of image quantization levels (32-level for GOLM1 and RND1, and 8-level for SETD7). Therefore, results for the identified image quantization levels are reported in this study.

Behavior of the generated multivariable regression model to predict gene expression signatures.
Fig. 4  Behavior of the generated multivariable regression model to predict gene expression signatures.

The association between regression model outcome and gene expression signatures is improving for training and validation. Even though the correlation difference between training and validation is increasing, the model is getting more complex. Therefore, the final regression models were generated with key variables selected after analysis of the model performances.

Spearman correlation coefficients of GOLM1, SETD7, and RND 1 genes improved from 0.38 to 0.79, 0.3 to 0.53, and 0.33 to 0.68 in the training cohort and from 0.37 to 0.56, 0.3 to 0.53, and 0.32 to 0.64 in the validation cohort with an increasing number of variables in the regression model. The results in the r-value for training and validation of the final regression model were seen in Table 4. The r -value of the validation were all greater than 0.5, and the RND model was the best (r =0.67).

Table 4

The models resulting in an r-value for training and validation cohorts.

ModelTraining
Validation
Spearman correlation coefficientp valueSpearman correlation coefficientp value
GOLM10.60<0.0010.500.06
STED70.53<0.0010.530.037
RND10.68<0.0010.670.046

Discussion

We investigated the potential value of radiomics analysis to determine the correlation between imaging features from structural MRI data and expression signatures of GOLM1, SETD7, and RND1 genes associated with HCC recurrence by building multivariable regression models using a machine learning method. The quantitative texture features from T2-weighted MRI showed a positive correlation with gene expression levels in early-stage HCC patients. These results demonstrated that preoperative MRI data can reveal gene expression signatures noninvasively, which could be utilized to predict recurrence risk for HCC patients instead of using the invasive and expensive genomic tests.

Detection of gene expression profiles and characteristics of the malignant tissues has great potential to reveal complex mechanisms of cancer biology and improve the prediction of disease recurrence or other types of outcomes for each patient. The previous genetic studies have demonstrated the potential values of GOLM1, SETD7, and RND1 genes in the evaluation of tumor growth, metastasis, and prognosis for HCC patients.11–13 Preoperative biopsy is a gold-standard clinical method to measure gene expression profiles. However, the costly and invasive biopsy cannot represent the whole tumor structure.22,31,32 Therefore, the association between noninvasive imaging characteristics and gene expression profiles becomes an active research topic for cancer studies during recent years.33,34

Several studies investigated the relationship between quantitative features of noninvasive imaging data and gene expression levels associated with the prognosis of HCC disease. Taouli et al.35 examined the correlation between 13 previously reported HCC gene signatures and characteristic features (11 qualitative and 4 quantitative) of noninvasive imaging (computed tomography or MRI) data. The results from 38 HCC patients showed a correlation between phenotypic imaging parameters and gene signatures of aggressive HCC disease. Hectors et al.36 investigated the heterogeneity of HCC tumors utilizing multiparametric MRI data and also evaluated the correlations between quantitative MRI parameters and gene expression signatures. Despite the observed significant correlation between expression levels of GLUL, FGFR4, EPCAM, and KRT19 and advanced MRI method parameters. Reproducibility is one of the main challenges that limits the generalization of the results. Xia et al.23 identified radiomics features extracted from contrast-enhanced CT data for the prognosis of HCC patients. Eight radiomics features selected from imaging data were significantly correlated with genes associated with HCC prognosis in 38 patients. Besides, they also found two features with significant correlation with OS. However, the authors stated the low number of patient cohorts limited the study in reproducibility.

In this study, we explored the value of quantitative MRI features for noninvasive extraction of gene information via correlating the MR features with gene expression levels of GOLM1, SET7, and RND1 associated with HCC recurrence using multivariable regression models. The quantitative T2-weighted MRI radiomics features were extracted and evaluated by performing a multi-step procedure. The developed regression models with the identified important variables showed a significant correlation for SETD7 and RND1 gene expression levels. Meanwhile, the GOLM1 model had a moderate correlation with a p-value closer to the significance level. The obtained results demonstrated the potential of quantitative analysis of conventional MRI data as a noninvasive tool for evaluation of the recurrence of the disease in HCC patients.

There are several limitations in our study. First, single-modality MRI data were analyzed to characterize tumor structure to predict gene expression levels. The correlation of image features with gene expression levels might be improved utilizing multiparametric MRI data. Second, we only analyzed three genes for predicting the risk of HCC recurrence in early-stage HCC patients. Even though these three genes have been applied in clinical practice as a prediction for the risk of HCC recurrence. The radiomics analysis for early prediction of the recurrence risk could benefit from an improved number of gene types. Third, we did not analyze the co-expression of the genes due to the small set of patient cohorts. Further studies including a larger number of the patient cohort with specialized patient selection procedure will improve interpretation of the imaging data for associating tumor characteristics and gene expression signatures.

In conclusion, the generated multivariable regression models obtained a good correlation with gene expression signatures associated with the recurrence of HCC in early-stage HCC patients. Our results indicate that quantitative analysis of MRI images had the potential to serve as a noninvasive approach for predicting gene expression levels by interpreting the tumor characteristics for early-stage HCC patients.

Supporting information

Supplementary Table 1

Radiomic features computed from computed tomography images.

(DOCX)

Supplementary Table 2

Key features selected extracted from MRI data quantized with 16 levels.

(DOCX)

Supplementary Table 3

Key features selected extracted from MRI data quantized with 32 levels.

(DOCX)

Supplementary Table 4

Key features selected extracted from MRI data quantized with 64 levels.

(DOCX)

Supplementary Table 5

Correlation between features and gene expression.

(DOCX)

Abbreviations

FOS: 

first-order statistics

GLCM: 

gray-level co-occurrence matrix

GLRLM: 

gray-level run-length matrix

GLSZM: 

gray-level size-zone matrix

GOLM1: 

Golgi membrane protein 1

HCC: 

hepatocellular carcinoma

LBP: 

local binary patterns

MRI: 

magnetic resonance imaging

NGTDM: 

neighborhood gray-tone difference matrix

OS: 

overall survival

RND1: 

Rho family GTPase 1

ROI: 

region of interest

SETD7: 

SET domain containing 7

Declarations

Data sharing statement

All data generated or analyzed in this study are available from the corresponding author for the reasonable request.

Funding

This study was supported by the National Key Research and Development Program of China (No. 2016YFC0107101 and No. 2016YFC0107109).

Conflict of interest

The authors have no conflict of interests related to this publication.

Authors’ contributions

Study concept and design (XL, LC), acquisition of the data (CmL, XlH), analysis and interpretation of the data (CL, XfH), drafting of the manuscript (XL), critical revision of the manuscript for important intellectual content (JW), and study supervision (LT). All the authors read and approved the final manuscript.

References

  1. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin 2019;69(1):7-34 View Article PubMed/NCBI
  2. American Cancer Society. Global Cancer Facts & Figures 2019. Available from: https://www.cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/annual-cancer-facts-and-figures/2019/cancer-facts-and-figures-2019.pdf View Article PubMed/NCBI
  3. Portolani N, Coniglio A, Ghidoni S, Giovanelli M, Benetti A, Tiberio GA, et al. Early and late recurrence after liver resection for hepatocellular carcinoma: prognostic and therapeutic implications. Ann Surg 2006;243(2):229-235 View Article PubMed/NCBI
  4. Xu XF, Xing H, Han J, Li ZL, Lau WY, Zhou YH, et al. Risk factors, patterns, and outcomes of late recurrence after liver resection for hepatocellular carcinoma: a multicenter study from China. JAMA Surg 2019;154(3):209-217 View Article PubMed/NCBI
  5. Woo HG, Park ES, Thorgeirsson SS, Kim YJ. Exploring genomic profiles of hepatocellular carcinoma. Mol Carcinog 2011;50(4):235-243 View Article PubMed/NCBI
  6. Hoshida Y, Villanueva A, Kobayashi M, Peix J, Chiang DY, Camargo A, et al. Gene expression in fixed tissues and outcome in hepatocellular carcinoma. N Engl J Med 2008;359(19):1995-2004 View Article PubMed/NCBI
  7. Skawran B, Steinemann D, Weigmann A, Flemming P, Becker T, Flik J, et al. Gene expression profiling in hepatocellular carcinoma: upregulation of genes in amplified chromosome regions. Mod Pathol 2008;21(5):505-516 View Article PubMed/NCBI
  8. Villanueva A, Hoshida Y, Battiston C, Tovar V, Sia D, Alsinet C, et al. Combining clinical, pathology, and gene expression data to predict recurrence of hepatocellular carcinoma. Gastroenterology 2011;140(5):1501-1512.e2 View Article PubMed/NCBI
  9. Li N, Li L, Chen Y. The identification of core gene expression signature in hepatocellular carcinoma. Oxid Med Cell Longev 2018;2018:3478305 View Article PubMed/NCBI
  10. Itzel T, Spang R, Maass T, Munker S, Roessler S, Ebert MP, et al. Random gene sets in predicting survival of patients with hepatocellular carcinoma. J Mol Med (Berl) 2019;97(6):879-888 View Article PubMed/NCBI
  11. Komatsu H, Iguchi T, Masuda T, Hirata H, Ueda M, Kidogami S, et al. Attenuated RND1 expression confers malignant phenotype and predicts poor prognosis in hepatocellular carcinoma. Ann Surg Oncol 2017;24(3):850-859 View Article PubMed/NCBI
  12. Ye QH, Zhu WW, Zhang JB, Qin Y, Lu M, Lin GL, et al. GOLM1 modulates EGFR/RTK cell-surface recycling to drive hepatocellular carcinoma metastasis. Cancer Cell 2016;30(3):444-458 View Article PubMed/NCBI
  13. Chen Y, Yang S, Hu J, Yu C, He M, Cai Z. Increased expression of SETD7 promotes cell proliferation by regulating cell cycle and indicates poor prognosis in hepatocellular carcinoma. PLoS One 2016;11(5):e0154939 View Article PubMed/NCBI
  14. Li Y, Eresen A, Lu Y, Yang J, Shangguan J, Velichko Y, et al. Radiomics signature for the preoperative assessment of stage in advanced colon cancer. Am J Cancer Res 2019;9(7):1429-38 View Article PubMed/NCBI
  15. Banerjee S, Wang DS, Kim HJ, Sirlin CB, Chan MG, Korn RL, et al. A computed tomography radiogenomic biomarker predicts microvascular invasion and clinical outcomes in hepatocellular carcinoma. Hepatology 2015;62(3):792-800 View Article PubMed/NCBI
  16. Eresen A, Yang J, Shangguan J, Li Y, Hu S, Sun C, et al. MRI radiomics for early prediction of response to vaccine therapy in a transgenic mouse model of pancreatic ductal adenocarcinoma. J Transl Med 2020;18(1):61 View Article PubMed/NCBI
  17. Jeong WK, Jamshidi N, Felker ER, Raman SS, Lu DS. Radiomics and radiogenomics of primary liver cancers. Clin Mol Hepatol 2019;25(1):21-29 View Article PubMed/NCBI
  18. Książek W, Abdar M, Acharya UR, Pławiak P. A novel machine learning approach for early detection of hepatocellular carcinoma patients. Cognitive Systems Research 2019;54:116-127 View Article PubMed/NCBI
  19. Gevaert O, Mitchell LA, Achrol AS, Xu J, Echegaray S, Steinberg GK, et al. Glioblastoma multiforme: exploratory radiogenomic analysis by using quantitative image features. Radiology 2014;273(1):168-174 View Article PubMed/NCBI
  20. Guo W, Li H, Zhu Y, Lan L, Yang S, Drukker K, et al. Prediction of clinical phenotypes in invasive breast carcinomas from the integration of radiomics and genomics data. J Med Imaging (Bellingham) 2015;2(4):041007 View Article PubMed/NCBI
  21. Li S, Han H, Sui D, Hao A, Qin H. A novel radiogenomics framework for genomic and image feature correlation using deep learning. , ; View Article PubMed/NCBI
  22. Lo Gullo R, Daimiel I, Morris EA, Pinker K. Combining molecular and imaging metrics in cancer: radiogenomics. Insights Imaging 2020;11(1):1 View Article PubMed/NCBI
  23. Xia W, Chen Y, Zhang R, Yan Z, Zhou X, Zhang B, et al. Radiogenomics of hepatocellular carcinoma: multiregion analysis-based identification of prognostic imaging biomarkers by integrating gene data-a preliminary study. Phys Med Biol 2018;63(3):035044 View Article PubMed/NCBI
  24. Hectors SJ, Wagner M, Bane O, Besa C, Lewis S, Remark R, et al. Quantification of hepatocellular carcinoma heterogeneity with multiparametric magnetic resonance imaging. Sci Rep 2017;7(1):2452 View Article PubMed/NCBI
  25. Lehmann TM, Gönner C, Spitzer K. Survey: interpolation methods in medical image processing. IEEE Trans Med Imaging 1999;18(11):1049-1075 View Article PubMed/NCBI
  26. Zwanenburg A, Leger S, Vallières M, Löck S. Image biomarker standardisation initiative. Available from: https://ui.adsabs.harvard.edu/abs/2016arXiv161207003Z View Article PubMed/NCBI
  27. Imamoglu N, Yagci A, Eresen A, Özbayoglu A. Intelligent Engineering Systems through Artificial Neural Networks. ASME Press; 2009, 447-454 View Article PubMed/NCBI
  28. Amadasun M, King R. Textural features corresponding to textural properties. IEEE Transactions on Systems, Man, and Cybernetics 1989;19(5):1264-1274 View Article PubMed/NCBI
  29. Thibault G, Fertil B, Navarro CL, Pereira S, Cau P, Lévy N, et al. Texture indexes and gray level size zone matrix. Application to cell nuclei classification. 10th International Conference on Pattern Recognition and Information Processing, PRIP 2009; Minsk, Belarus. View Article PubMed/NCBI
  30. Kononenko I, Šimec E, Robnik-Šikonja M. Overcoming the Myopia of Inductive Learning Algorithms with RELIEFF. Applied Intelligence 1997;7(1):39-55 View Article PubMed/NCBI
  31. Tam AL, Lim HJ, Wistuba II, Tamrazi A, Kuo MD, Ziv E, et al. Image-guided biopsy in the era of personalized cancer care: proceedings from the society of interventional radiology research consensus panel. J Vasc Interv Radiol 2016;27(1):8-19 View Article PubMed/NCBI
  32. Pritzker KPH, Nieminen HJ. Needle biopsy adequacy in the era of precision medicine and value-based health care. Arch Pathol Lab Med 2019;143(11):1399-1415 View Article PubMed/NCBI
  33. Hectors SJ, Cherny M, Yadav KK, Beksaç AT, Thulasidass H, Lewis S, et al. Radiomics features measured with multiparametric magnetic resonance imaging predict prostate cancer aggressiveness. J Urol 2019;202(3):498-505 View Article PubMed/NCBI
  34. Wang T, Gong J, Duan HH, Wang LJ, Ye XD, Nie SD. Correlation between CT based radiomics features and gene expression data in non-small cell lung cancer. J Xray Sci Technol 2019;27(5):773-803 View Article PubMed/NCBI
  35. Taouli B, Hoshida Y, Kakite S, Chen X, Tan PS, Sun X, et al. Imaging-based surrogate markers of transcriptome subclasses and signatures in hepatocellular carcinoma: preliminary results. Eur Radiol 2017;27(11):4472-4481 View Article PubMed/NCBI
  36. Hectors SJ, Wagner M, Bane O, Besa C, Lewis S, Remark R, et al. Quantification of hepatocellular carcinoma heterogeneity with multiparametric magnetic resonance imaging. Sci Rep 2017;7(1):2452 View Article PubMed/NCBI