Intratumoral and Peritumoral Radiomics Based on Contrast-enhanced MRI for Preoperatively Predicting Treatment Response of Transarterial Chemoembolization in Hepatocellular Carcinoma

DOI: https://doi.org/10.21203/rs.3.rs-2190531/v1

Abstract

Background

Noninvasive and precise methods to estimate treatment response and identify hepatocellular carcinoma (HCC) patients who could benefit from transarterial chemoembolization (TACE) are urgently required. The present study aimed to investigate the ability of intratumoral and peritumoral radiomics based on contrast-enhanced magnetic resonance imaging (CE-MRI) to preoperatively predict tumor response to TACE in HCC patients.

Methods

This retrospective study involved 138 HCC patients (objective response, n = 73 vs. non-response, n = 65) who were divided into the training (n = 96) and validation (n = 42) cohorts. Total 1206 radiomics features were extracted from arterial, venous, and delayed phases images. Radiomics models on intratumoral region (TR) and peritumoral region (PTR) (3 mm, 5 mm, and 10 mm) were established using logistic regression. Three integrated radiomics models, including intratumoral and peritumoral region (T-PTR) (3 mm), T-PTR (5 mm), and T-PTR (10 mm) models, were constructed by using TR and PTR radiomics scores. A clinical-radiological model and a combined model incorporating the optimal radiomics score and selected clinical-radiological predictors were constructed, and the combined model was presented as a nomogram. The discrimination, calibration, and clinical utilities were evaluated by receiver operating characteristic curve, calibration curve, and decision curve analysis, respectively.

Results

The (T-PTR) (3 mm) radiomics model demonstrated the best performance among all radiomics models with an area under the curve (AUC) of 0.911 (95% confidence interval(CI), 0.825–0.975) in the validation cohort. The (T-PTR) (3 mm) radiomics score, alkaline phosphatase, tumor size, and satellite nodule were combined to construct a combined nomogram. The combined nomogram outperformed the clinical-radiological model with the AUCs of 0.918 (95%CI, 0.831–0.986) and 0.782 (95%CI, 0.660–0.902) and achieved good calibration capability and clinical utility.

Conclusions

CE-MRI-based intratumoral and peritumoral radiomics approach can provide an effective tool for the precise and individualized estimation of treatment response for HCC patients treated with TACE.

Introduction

Hepatocellular carcinoma (HCC) is the most common liver malignancy and the third leading cause of death among various cancers[1], with more than 850,000 associated deaths reported annually worldwide[2]. Liver transplantation and resection are the curative therapies for HCC, and surgical excision is considered as the preferred treatment for HCC patients at early stage[3]. Unfortunately, the majority of HCC patients are not suitable for curative resection at the time of diagnosis because of poor liver function, multifocal disease, vascular involvement, and extrahepatic spread[4]. Transarterial chemoembolization (TACE) is widely used as a bridge to liver transplantation, or as the standard treatment for patients with intermediate HCC[5]. Nevertheless, the therapeutic efficacy of TACE varies greatly between patients and patients because the biological behavior of HCC is highly heterogeneous[6]. Several studies have evidenced that the overall response rates following TACE range from 15–85% and the cumulative rates of local tumor progression at 1, 3, and 5 years are 33%, 52%, and 73%, respectively[7, 8]. Therefore, it is beneficial to preoperatively estimate tumor response to TACE treatment and determine the therapeutic efficacy to aid in guiding subsequent therapeutic strategies to improve overall survival (OS) of HCC patients.

Magnetic resonance imaging (MRI)-based evaluations that are noninvasive and repeatable can be used to preoperatively assess the tumor response, which take an advantage of depicting more soft-tissue characteristics than computed tomography (CT) and avoiding ionizing radiation[9]. Regarding the pretherapeutic imaging features of HCC, several scholars have reported that larger lesion diameter, irregular margin, arterial peritumoral enhancement, satellite nodule, and quantitative functional parameters of diffusion weighted imaging (DWI), intravoxel incoherent motion (IVIM), diffusion kurtosis imaging (DKI), and dynamic contrast enhanced magnetic resonance imaging (DCE-MRI), are found to be associated with therapeutic efficacy of TACE treatment[915]. Although these imaging characteristics are encouraging, they are not sufficient for individual evaluations of tumor response to TACE, and the ability to predict TACE efficacy in HCC is limited when a high degree of tumor heterogeneity.

Radiomics is an emerging methodology, with the goal of going beyond size or human-eye based semantic descriptors of tumors, to enable the non-invasive extraction of high-throughput quantitative data from medical images and to explore their correlation with clinical outcomes[16]. This approach would capture the three-dimensional (3D) information of the entire tumor and has been proved effective in characterizing the heterogeneity of the tumor by acting as a whole tumor virtual biopsy[17, 18]. Numerous studies have demonstrated that radiomics-based models effectively identify the diagnosis and pathological characteristics or predict therapeutic efficacy and prognosis of cancer patients for clinical decision-making[1922]. Recently, there has been increasing interest in evaluating radiomics patterns of the region surrounding the visible tumor[20, 23, 24]. Recurrence or metastasis of HCC is mainly intrahepatic, indicating that the peritumoral liver tissue may be a favorable soil for the spreading hepatoma cells[25]. Several scholars have reported that HCC patients with microvascular invasion (MVI), epithelial cell adhesion molecule (EpCAM), programmed death ligand 1 (PD-L1) expression, and higher CD68 + cell density in peritumoral tissues have a significantly higher risk of recurrence or metastasis and cancer-related death[2629]; thus, peritumoral tissues might have valuable predictive information of HCC prognosis. Several recent studies have focused on peritumoral radiomics for predicting MVI, early recurrence (ER), recurrence-free survival (RFS) of HCC patients after resection or TACE[20, 23, 24]; however, the value of intratumoral and peritumoral radiomics based on MRI in predicting treatment response of HCC after TACE remains unknown.

Therefore, the present study aimed to determine whether radiomics assessment of HCC peritumoral regions based on contrast-enhanced MR (CE-MR) images could provide valuable information about TACE response and enhance the ability of intratumoral radiomics to predict treatment efficacy of TACE in patients with HCC.

Materials And Methods

Patients

This retrospective study was approved by the Institutional Review Board of the First Affiliated Hospital of Dalian Medical University and the requirement for informed consent was waived due to the retrospective nature of the study.

From April 2008 to February 2022, 343 consecutive patients with HCC who underwent CE-MRI examination within two weeks before conventional TACE at our institution were recruited. HCCs were diagnosed by pathology or with reference to American Association for the Study of Liver Disease (AASLD) guidelines[30]. The exclusion criteria were: (1) previous oncological treatment, including liver resection, radiofrequency ablation (RFA), or chemotherapy (n = 32); (2) diffuse or infiltrative HCCs or the largest lesion size < 1cm (n = 16); (3) extrahepatic metastasis or portal vein occlusion (n = 10); (4) unavailability of hepatic-arterial CE-MR imaging within 3 months after TACE treatment (n = 145); (5) poor image quality (n = 2). Figure 1 shows the flowchart of patient recruitment, and 138 patients were enrolled in this study and randomly divided into a training cohort (n = 96) and a validation cohort (n = 42) at a ratio of 7:3.

Pretherapeutic clinical characteristics including age, gender, history of hepatitis B or C, alpha-fetoprotein (AFP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), γ-glutamyltranspeptadase (GGT), alkaline phosphatase (ALP), total bilirubin (TBIL), albumin (ALB), platelet count (PLT), prothrombin time (PT), Child-Pugh class, Eastern Cooperative Oncology Group (ECOG) performance status, and Barcelona Clinic Liver Cancer (BCLC) stage for each patient were retrospectively collected within 1 week before TACE.

MRI Protocol

All MRI examinations were performed with 1.5T or 3.0T MR scanner (Signa, HDXT, GE Healthcare) with a phased-array 8-channel sensitivity encoding abdominal coil. The CE-MRI examination was performed using the liver acquisition with volume acceleration (LAVA) protocol with fat-suppressed T1-weighted 3D fast-spoiled gradient-recalled echo sequence. The contrast enhanced images consisted of arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) images, which were obtained at 40 s, 70 s, and 90 s, respectively, after the start of contrast injection of Gd-diethylenetriamine pentaacetic acid (Gd-DTPA) (Bayer Schering Pharma AG, Germany) at a patient weight-dependent dose of 0.1 mmol/kg with an injection rate of 2.5 mL/s through a median cubital vein. Other MRI sequences were also performed, including in- and opposed-phase fast-spoiled gradient-recalled echo T1-weighted imaging (T1WI) and fat-suppressed fast spin-echo T2-weighted imaging (T2WI). The detailed scanning parameters are shown in Supplementary Data S1.

Image Analysis

Two radiologists (reader 1, Y.Z. and reader 2, N.W., with 8-year and 7-year experience in abdominal MRI) independently analyzed pretherapeutic MR images, and they were aware of the diagnosis of HCC but blinded to clinical information and imaging report. The radiologists evaluated the following imaging traits for each patient: (1) tumor size; (2) tumor number; (3) tumor margin; (4) intratumoral necrosis; (5) intratumoral hemorrhage; (6) intratumoral fat; (7) tumor encapsulation; (8) arterial peritumoral enhancement; (9) satellite nodule; (10) internal arteries; (11) radiological cirrhosis. Tumor diameter was recorded as mean value, and any discrepancy in imaging feature assessment was resolved by means of reevaluation by another senior radiologist (J.H.L., with 20-year experience in abdominal MRI). Detailed definitions of these imaging features are described in Supplementary Data S2.

TACE Procedure and Treatment Response Assessment

All conventional TACE procedures were carried out by interventional radiologists with no less than 10 years of clinical experience. The Seldinger technique was used to puncture the right femoral artery, and a 5F catheter (RH catheter, Terumo, Tokyo, Japan) was inserted for celiac and hepatic artery angiography in evaluation of the vascular anatomy and tumor burden. After superselective catheterization of hepatic artery feeding the tumor, an emulsion mixture of lobaplatin, raltitrexed, and Lipiodol was injected using a 2.7F microcatheter (Progreat microcatheter, Hengrui Pharmaceutical Co. Ltd., Lianyungang, China), followed by embolization of the tumor-feeding artery using blank microsphere (100–300 µm) or gelatin sponge particles (1–2 mm). The TACE procedure was stopped when tumor staining completely disappeared and regional arterial blood flow stopped.

The modified Response Evaluation Criteria in Solid Tumors (mRECIST 1.1) criteria was utilized to assess the tumor response in patients with HCC based on pre- and post-therapeutic arterial MR images. Tumor response was classified into four categories according to the mRECIST system as follows: complete response (CR), partial response (PR), stable disease (SD), and progression disease (PD)[31]. In the present study, all patients were divided into the objective response (OR) group (CR and PR patients) and the non-response (NR) group (SD and PD patients).

Tumor Segmentation and Radiomics Feature Extraction

Pretherapeutic CE-MR images were exported from the picture archiving and communication system (PACS) and then used for tumor segmentation and radiomics feature extraction. The AK software (Artificial Intelligence Kit, Version 3.2.5, GE Healthcare) was used to process the images before segmenting the tumor. AP, PVP, and DP images were resampled to a uniform voxel size of 1 × 1 × 1 mm via linear interpolation algorithm to standardize the voxel spacing[22]. Intensity normalization of images was performed to correct the scanner effect. Tumor segmentation was performed by manually delineating the region of interest (ROI) along the tumor contour on each axial slice of AP, PVP, and DP images using an open-source software (ITK-SNAP, version 3.6.0, http://www.itksnap.org/). The ROI was required to include capsule surrounding the tumor and to exclude tumor surrounding vessels, and then every ROI was automatically merged into volume of interest (VOI). Notably, in terms of multifocal HCCs, the largest nodule was selected as the delineated lesion. To capture radiomics features from the tumor periphery, the VOIs of peritumoral region (PTR) were generated by automatically expanding 3 mm, 5 mm, and 10 mm from the lesion border using AK software. If the ROI was beyond the parenchyma of the liver after the expansion, the portion beyond the parenchyma was removed manually.

Radiomics features were extracted using AK software for intratumoral region (TR), PTR (3 mm), PTR (5 mm), and PTR (10 mm). A total of 1206 radiomics features were extracted for each VOI (based on AP, PVP, and DP images). The extracted radiomics features included: 42 histograms, 15 form factors, 10 Haralick features, 144 gray level co-occurrence matrix (GLCM) with an offset of 1/4/7, 180 gray level run length matrix (GLRLM) with an offset of 1/4/7, and 11 grey-level zone size matrix (GLZSM). Details of radiomics features are listed in Supplementary Data S3. Z-score normalization of radiomics features was performed to reduce the bias caused by different dimensions. The workflow of the radiomics analysis is depicted in Fig. 2.

Inter- and intraclass correlation coefficients (ICCs) were used to assess the reproducibility of each radiomics feature extracted from 30 randomly chosen patients. To assess the interobserver reproducibility, the ROI delineation was performed by two abdominal radiologists (readers 1 and 2, Y.Z. and N.W.) independently who were blinded to all patients’ information. To evaluate the intraobserver reproducibility, reader 1 repeated the ROI delineation and feature extraction at a 1-month interval. Finally, reader 1 completed the remaining image segmentation and feature extraction.

Feature Selection and Radiomics Model Construction

To remove potentially redundant features and decrease data dimensions, we followed a three-step procedure to identify the most predictive radiomics features. First, the radiomics features with ICCs of both interobserver and intraobserver greater than 0.80 indicated satisfactory reproducibility and high robustness and were selected for further analysis. Second, the redundant features with a correlation coefficient higher than 0.90 were eliminated following by the Spearman’s rank correlation test. Finally, the gradient boosting decision tree (GBDT) algorithm was applied to determine the top-ranked and most valuable features for predicting the tumor response. GBDT was proposed as a tree-based algorithm based on a greedy strategy (called gradient boosting) that evaluates the importance of a texture feature through the time it used as branching point for the tree[19]. After those steps, TR, PTR (3 mm), PTR (5 mm), and PTR (10 mm) radiomics models were separately established using logistic regression algorithm with 5-fold cross-validation. The radiomics score (Rad-score) was calculated for each patient using a linear combination of the selected features that were weighted by their respective coefficients. Finally, three integrated radiomics models, including intratumoral and peritumoral region (T-PTR) (3 mm), T-PTR (5 mm), and T-PTR (10 mm) radiomics models, were generated by logistic regression using TR rad-score and corresponding PTR rad-score. The optimal radiomics model with the highest area under the curve (AUC) was selected for further analysis.

Clinical-radiological Model Construction

Univariate analysis was used to identify significant variables among clinical-radiological characteristics associated with treatment response of HCC (P < 0.05). Multivariate logistic regression analysis was performed to identify the independent risk factors for predicting tumor response. Odds ratio and 95% confidence interval (CI) were calculated for each risk factor. The clinical-radiological model was constructed using the chosen independent risk factors.

Combined Model Building and Nomogram Construction

A combined model integrating clinical-radiological risk factors and the optimal rad-score was constructed using the proposed logistic regression method. To provide an individual predictive graphical presentation, the combined model was presented as a nomogram. The nomogram could help calculate the predicted probability of NR for each individual patient.

Statistical Analysis

The Student’s t-test or Mann-Whitney U-test were used to compare the continuous variables between the OR and NR groups as appropriate. The chi-squared test or Fisher’s exact test were used to assess the categorical variables as appropriate. Interobserver agreement of the radiological feature evaluation was qualified by Cohen kappa coefficient. Kappa values of 0.81–1.00 indicated excellent agreement, 0.61–0.80 signified substantial agreement, and 0.41–0.60 denoted moderate agreement. Receiver operating characteristic curve (ROC) analysis was performed to evaluate the performance of each predictive model. The AUC, accuracy, sensitivity, and specificity were calculated. We compared the predictive performance between different models using the Delong’s test. Calibration curves were drawn to evaluate the degree of deviation between the predictions and actual outcomes with the Hosmer-Lemeshow test. Decision curve analysis (DCA) was performed to validate the clinical utility of the nomogram. All statistical analyses for the present study were performed with R software (version 3.6.1, http://www.R-project.org). P < 0.05 was considered statistically significant.

Results

Patient Characteristics

In total, 138 patients (mean age, 60.24 ± 8.91 years; range, 40–83 years; 115 male) were enrolled in this study. Among the 138 patients with HCC, 6 patients were determined by pathology and 132 patients were identified by specific imaging features according to the AASLD guidelines. On the basis of the mRECIST criteria, the patients for CR, PR, SD, and PD were 13 (9.4%), 60 (43.5%), 51 (37.0%), and 14 (10.1%), respectively. All patients underwent post-therapeutic MRI examination following by the initial TACE treatment, with a median interval of 44 days (range, 28–90 days) between the first TACE and the follow-up MRI examination. The clinical-radiological data in the OR and NR groups are summarized in Table 1. No significant difference was found in the clinical-radiological characteristics, except for tumor encapsulation between the training and validation cohorts. 

 Table 1. Patient clinical-radiological characteristics

Characteristics

Training cohort (n = 96)

 

Validation cohort (n = 42)

value

OR group

(n = 51)

NR group

(n = 45)

value

 

OR group

(n = 22)

NR group

(n = 20)

value

Age (years, mean ± SD)

60.29±8.41

60.42±10.25

0.947

 

60.18±9.02

59.75±7.31

0.866

0.820

Gender (n, [%])

 

 

0.410

 

 

 

0.890

1.000

Male

44(86.3)

36(80.0)

 

 

18(81.8)

17(85.0)

 

 

Female

7(13.7)

9(20.0)

 

 

4(18.2)

3(15.0)

 

 

History of hepatitis B or C (n, [%])

 

 

0.541

 

 

 

0.361

0.410

Positive

38(74.5)

31(68.9)

 

 

19(86.4)

14(70.0)

 

 

Negative

13(25.5)

14(31.1)

 

 

3(13.6)

6(30.0)

 

 

AFP (IU/ml) (n, [%])

 

 

0.050

 

 

 

0.231

0.523

≤ 400

40(78.4)

27(60.0)

 

 

16(72.7)

11(55.0)

 

 

> 400

11(21.6)

18(40.0)

 

 

6(27.3)

9(45.0)

 

 

ALT (U/L) (n, [%])

 

 

0.521

 

 

 

0.827

0.904

≤ 50

36(70.6)

29(64.4)

 

 

15(68.2)

13(65.0)

 

 

> 50

15(29.4)

16(35.6)

 

 

7(31.8)

7(35.0)

 

 

AST (U/L) (n, [%])

 

 

0.024

 

 

 

0.187

0.302

≤ 40

31(60.8)

17(37.8)

 

 

11(50.0)

6(30.0)

 

 

> 40

20(39.2)

28(62.2)

 

 

11(50.0)

14(70.0)

 

 

GGT (U/L) (n, [%])

 

 

0.004

 

 

 

0.016

0.639

≤ 60

26(51.0)

10(22.2)

 

 

11(50.0)

3(15.0)

 

 

> 60

25(49.0)

35(77.8)

 

 

11(50.0)

17(85.0)

 

 

ALP (U/L) (n, [%])

 

 

<0.001

 

 

 

0.005

0.313

≤ 125

46(90.2)

26(57.8)

 

 

19(86.4)

9(45.0)

 

 

> 125

5(9.8)

19(42.2)

 

 

3(13.6)

11(55.0)

 

 

TBIL (umol/L) (n, [%])

 

 

0.650

 

 

 

0.746

0.177

≤ 19

34(66.7)

28(62.2)

 

 

11(50.0)

11(55.0)

 

 

> 19

17(33.3)

17(37.8)

 

 

11(50.0)

9(45.0)

 

 

ALB (g/L) (n, [%])

 

 

0.897

 

 

 

0.789

0.922

< 40

29(56.9)

25(55.6)

 

 

13(59.1)

11(55.0)

 

 

≥ 40

22(43.1)

20(44.4)

 

 

9(40.9)

9(45.0)

 

 

PLT (×109/L) (n, [%])

 

 

0.031

 

 

 

0.516

0.783

< 125

27(52.9)

14(31.1)

 

 

11(50.0)

8(40.0)

 

 

    ≥ 125

24(47.1)

31(68.9)

 

 

11(50.0)

12(60.0)

 

 

PT (s) (n, [%])

 

 

0.339

 

 

 

0.569

0.192

≤ 13

34(66.7)

34(75.6)

 

 

14(63.6)

11(55.0)

 

 

> 13

17(33.3)

11(24.4)

 

 

8(36.4)

9(45.0)

 

 

Child-Pugh class (n, [%])

 

 

0.585

 

 

 

0.379

0.078

A

44(86.3)

37(82.2)

 

 

17(77.3)

13(65.0)

 

 

B

7(13.7)

8(17.8)

 

 

5(22.7)

7(35.0)

 

 

ECOG performance status (n, [%])

 

 

0.060

 

 

 

0.095

0.084

0

49(96.1)

37(82.2)

 

 

20(90.9)

13(65.0)

 

 

1

2(3.9)

8(17.8)

 

 

2(9.1)

7(35.0)

 

 

BCLC stage (n, [%])

 

 

0.003

 

 

 

0.076

0.480

A

35(68.6)

18(40.0)

 

 

13(59.1)

8(40.0)

 

 

B

13(25.5)

12(26.7)

 

 

7(31.8)

3(15.0)

 

 

C

3(5.9)

15(33.3)

 

 

2(9.1)

9(45.0)

 

 


Table 1. Patient clinical-radiological characteristics (continued)

Characteristics

Training cohort (n = 96)

 

Validation cohort (n = 42)

value

OR group

(n = 51)

NR group

(n = 45)

value

 

OR group

(n = 22)

NR group

(n = 20)

value

Tumor size (n, [%])

 

 

<0.001

 

 

 

0.001

0.936

≤ 5 cm

36(70.6)

15(33.3)

 

 

17(77.3)

5(25.0)

 

 

> 5 cm

15(29.4)

30(66.7)

 

 

5(22.7)

15(75.0)

 

 

Tumor number (n, [%])

 

 

0.664

 

 

 

0.275

1.000

Unifocal

35(68.6)

29(64.4)

 

 

13(59.1)

15(75.0)

 

 

Multifocal

16(31.4)

16(35.6)

 

 

9(40.9)

5(25.0)

 

 

Tumor margin (n, [%])

 

 

0.541

 

 

 

0.031

0.245

Smooth

38(74.5)

31(68.9)

 

 

17(77.3)

9(45.0)

 

 

Non-smooth

13(25.5)

14(31.1)

 

 

5(22.7)

11(55.0)

 

 

Intratumoral necrosis (n, [%])

 

 

0.005

 

 

 

0.129

0.833

Present

16(31.4)

27(60.0)

 

 

7(31.8)

11(55.0)

 

 

Absent

35(68.6)

18(40.0)

 

 

15(68.2)

9(45.0)

 

 

Intratumoral hemorrhage (n, [%])

 

 

0.030

 

 

 

0.060

0.611

Present

13(25.5)

21(46.7)

 

 

4(18.2)

9(45.0)

 

 

Absent

38(74.5)

24(53.3)

 

 

18(81.8)

11(55.0)

 

 

Intratumoral fat (n, [%])

 

 

0.112

 

 

 

0.872

0.826

Present

7(13.7)

12(26.7)

 

 

5(22.7)

4(20.0)

 

 

Absent

44(86.3)

33(73.3)

 

 

17(77.3)

16(80.0)

 

 

Tumor encapsulation (n, [%])

 

 

0.286

 

 

 

0.032

0.026

Present

39(76.5)

30(66.7)

 

 

15(68.2)

7(35.0)

 

 

Absent

12(23.5)

15(33.3)

 

 

7(31.8)

13(65.0)

 

 

Arterial peritumoral enhancement (n, [%])

 

 

0.050

 

 

 

0.118

0.846

Present

11(21.6)

18(40.0)

 

 

4(18.2)

8(40.0)

 

 

Absent

40(78.4)

27(60.0)

 

 

18(81.8)

12(60.0)

 

 

Satellite nodule (n, [%])

 

 

0.001

 

 

 

0.670

0.831

Present

1(2.0%)

11(24.4)

 

 

2(9.1)

2(10.0)

 

 

Absent

50(98.0)

34(75.6)

 

 

20(90.9)

18(90.0)

 

 

Internal arteries (n, [%])

 

 

0.046

 

 

 

0.031

0.464

Present

18(35.3)

25(55.6)

 

 

5(22.7)

11(55.0)

 

 

Absent

33(64.7)

20(44.4)

 

 

17(77.3)

9(45.0)

 

 

Radiological cirrhosis (n, [%])

 

 

0.030

 

 

 

0.372

0.719

Present

36(70.6)

22(48.9)

 

 

14(63.6)

10(50.0)

 

 

Absent

15(29.4)

23(51.1)

 

 

8(36.4)

10(50.0)

 

 

OR, objective response; NR, non-response; SD, standard deviation; AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyltranspeptadas; ALP, alkaline phosphatase; TBIL, total bilirubin; ALB, albumin; PLT, platelet count; PT, prothrombin time; ECOG, Eastern Cooperative Oncology Group; BCLC, Barcelona Clinic Liver Cancer. 

Radiomics Model Development and Evaluation

After inter- and intraobserver reproducibility analysis, the dimensions of feature spaces were 853 for VOITR, 424 for VOIPTR (3 mm), 452 for VOIPTR (5 mm), and 716 for VOIPTR (10 mm). Spearman’s rank correlation test allowed the selection of 100, 35, 43, and 90 features, respectively. GBDT revealed that the radiomics feature numbers ultimately consisted of 25, 14, 17, and 21 from the VOITR, VOIPTR (3 mm), VOIPTR (5 mm), and VOIPTR (10 mm), respectively, and were used for the radiomics model building (TR, PTR (3 mm), PTR (5 mm), and PTR (10 mm) models). Finally, three T-PTR radiomics models based on TR and PTR rad-scores were constructed. The calculation formulae for the rad-score are shown in Supplementary Data S4.

The radiomics models demonstrated favorable discrimination in the both cohorts (AUC: training cohort, 0.810–0.892; validation cohort, 0.793–0.911). The TR, PTR (3mm), PTR (5mm) radiomics models showed comparative performance with the AUCs of 0.820 (95% CI, 0.705–0.917), 0.823 (95% CI, 0.701–0.927), and 0.823 (95% CI, 0.711–0.924) in the validation cohort, and the predictive performance of these three models were slightly higher than that of the PTR (10 mm) model (AUC: 0.793; 95%CI: 0.673–0.903). Compared with TR and PTR radiomics models, the T-PTR radiomics models (T-PTR (3 mm), T-PTR (5 mm), and T-PTR (10 mm) models) performed better in predicting tumor response with the AUCs of 0.911 (95%CI, 0.825–0.975), 0.909 (95%CI, 0.817–0.982), and 0.895 (95%CI, 0.815–0.964) in the validation cohort; thus, we chose the T-PTR (3 mm) model as the best-performing radiomics model for further analysis. ROC curves and discriminative performance of the seven radiomics models in the two cohorts are shown in Supplementary Data S5 and Table 2

 

Model

AUC (95% CI)

Accuracy

Sensitivity

Specificity

P value

Table 2

Discrimination performance of different predictive models

Clinical-radiological model

TC

0.789 (0.709–0.863)

0.729

0.556

0.882

0.935

VC

0.782 (0.660–0.902)

0.714

0.600

0.818

TR model

TC

0.836 (0.763–0.903)

0.771

0.778

0.765

0.840

VC

0.820 (0.705–0.917)

0.714

0.750

0.682

PTR (3 mm) model

TC

0.817 (0.740–0.881)

0.750

0.889

0.627

0.940

VC

0.823 (0.701–0.927)

0.714

0.800

0.636

PTR (5 mm) model

TC

0.810 (0.739–0.880)

0.771

0.800

0.745

0.877

VC

0.823 (0.711–0.924)

0.714

0.900

0.545

PTR (10 mm) model

TC

0.852 (0.785–0.911)

0.760

0.822

0.706

0.468

VC

0.793 (0.673–0.903)

0.762

0.900

0.636

T-PTR (3 mm) model

TC

0.884 (0.821–0.936)

0.812

0.822

0.804

0.633

VC

0.911 (0.825–0.975)

0.810

0.800

0.818

T-PTR (5 mm) model

TC

0.883 (0.822–0.934)

0.823

0.733

0.902

0.661

VC

0.909 (0.817–0.982)

0.810

0.650

0.955

T-PTR (10 mm) model

TC

0.892 (0.834–0.942)

0.844

0.800

0.882

0.945

VC

0.895 (0.815–0.964)

0.786

0.750

0.818

Combined nomogram

TC

0.910 (0.854–0.958)

0.844

0.822

0.863

0.891

VC

0.918 (0.831–0.986)

0.857

0.800

0.909

TR, intratumoral region; PTR, peritumoral region; T-PTR, intratumoral and peritumoral region; TC, training cohort; VC, validation cohort; AUC, area under the curve; CI, confidence interval.

Clinical-radiological Model Development and Evaluation

Interobserver agreements on the radiological features were substantial to excellent (kappa-value range: 0.772–1.000). Univariate and multivariate analyses of clinical-radiological characteristics for predicting treatment response in the training cohort are shown in Table 3. The univariate analysis indicated that AST, GGT, ALP, PLT, ECOG, BCLC stage, tumor size, intratumoral necrosis, intratumoral hemorrhage, satellite nodule, internal arteries, and radiological cirrhosis were significant clinical-radiological factors for discriminating the OR and NR groups in the training cohort (all P < 0.05). The multivariate logistic regression analysis demonstrated that ALP (odd ratio = 5.744; 95% CI: 1.780–18.532; P = 0.003), tumor size (odd ratio = 3.005; 95% CI: 1.154–7.826; P = 0.024), and satellite nodule (odd ratio = 9.865; 95% CI: 1.101–88.370; P = 0.041) were the independent risk factors for predicting NR in HCC patients. The clinical-radiological model was built by incorporating these three variables. The clinical-radiological model yielded the AUCs of 0.789 (95% CI, 0.709–0.863) and 0.782 (95% CI, 0.660–0.902) in the two cohorts, respectively (shown in Table 2). 

 Table 3. Univariate and multivariate analyses of clinical-radiological characteristics for predicting treatment response

Variables

Univariate analysis

 

Multivariate analysis

Odd ratio (95% CI)

value

 

Odd ratio (95% CI)

value

AST

2.553 (1.120 - 5.820)

0.026

 

GGT

3.640 (1.492 - 8.880)

0.005

 

ALP

6.723 (2.246 - 20.122)

<0.001

 

5.744 (1.780 - 18.532)

0.003

PLT

2.491 (1.079 - 5.753)

0.033

 

ECOG

5.297 (1.062 - 26.428)

0.042

 

BCLC stage

2.716 (1.515 - 4.872)

<0.001

 

Tumor size

4.800 (2.023 - 11.392)

<0.001

 

3.005 (1.154 - 7.826)

0.024

Intratumoral necrosis

3.281 (1.417 - 7.600)

0.006

 

Intratumoral hemorrhage

2.558 (1.082 - 6.044)

0.032

 

Satellite nodule

16.176 (1.996 - 131.069)

0.009

 

9.865 (1.101 - 88.370)

0.041

Internal arteries

2.292 (1.007 - 5.213)

0.048

 

Radiological cirrhosis

0.399 (0.172 - 0.923)

0.032

 

The clinical-radiological characteristics with P value less than 0.05 in the univariate analysis are listed in the table. AST, aspartate aminotransferase; GGT, γ-glutamyltranspeptadase; ALP, alkaline phosphatase; PLT, platelet count; ECOG, Eastern Cooperative Oncology Group; BCLC, Barcelona Clinic Liver Cancer; CI, confidence interval.

Combined Model and Nomogram Development and Evaluation

The T-PTR (3 mm) rad-score, ALP, tumor size, and satellite nodule were considered as input variables for logistic regression to build the combined model. We chose a nomogram as the graphical representation of the best-performing combined model as shown in Fig. 3. The combined nomogram yielded an AUC, an accuracy, a sensitivity, and a specificity of 0.910 (95% CI, 0.854–0.958), 0.844, 0.822, and 0.863, respectively, for discriminating between the OR and NR groups in the training cohort and of 0.918 (95% CI, 0.831–0.986), 0.857, 0.800, and 0.909, respectively, in the validation cohort. The discriminating performance of the combined nomogram in the training cohort was significantly superior to that of the clinical-radiological model (AUC, 0.910 vs. 0.789, P = 0.028), whereas there was no significant difference between the models in the validation cohort (AUC, 0.918 vs. 0.782, P = 0.127). No significant differences were found between the T-PTR (3mm) radiomics model and combined nomogram (training cohort, P = 0.577; validation cohort, P = 0.918) and between the T-PTR (3mm) radiomics model and clinical-radiological model (training cohort, P = 0.094; validation cohort, P = 0.139). Table 2 summarizes the predictive performance of the constructed models. ROC curves for clinical-radiological model, T-PTR (3mm) radiomics model, and combined nomogram in the both cohorts are shown in Fig. 4A and 4B. The Delong’s test of different predictive models in the both cohorts is shown in Supplementary Data S6.

Calibration curves of the combined nomogram for the probability of treatment response demonstrated good agreements between prediction and observation in the training and validation cohorts (Fig. 5A, 5B). The Hosmer-Lemeshow test yielded non-significant results in the both cohorts (P = 0.386 and 0.343), which suggested a satisfying fit of the nomogram. The DCA indicated that the combined nomogram obtained more net benefits than the clinical-radiological model, T-PTR (3mm) radiomics model, and “treat-all” or “treat-none” strategies for most of the threshold probabilities in the training and validation cohorts (Fig. 6A, 6B).

Discussion

In the study, we analyzed various radiomics models for intratumoral, peritumoral, and intratumoral combined peritumoral derived from preoperative CE-MR images for response estimation after TACE. Our study confirms that the radiomics models combining intratumoral and peritumoral are better than the intratumoral radiomics model in predicting TACE response of HCC patients. Furthermore, we developed and validated a combined nomogram that integrated the clinical-radiological risk factors and the optimal T-PTR (3mm) rad-score for predicting HCC response to TACE, demonstrating satisfactory performance, calibration, and clinical utility. The proposed radiomics approach successfully predicted therapeutic efficacy of TACE and may facilitate individualized treatment decision-making for patients with HCC.

Patients with HCC receiving TACE have various treatment efficacy and clinical outcomes[7, 8]. Treatment response after first TACE course has been proved to be an important and robust prognostic factor for clinical outcomes, which also impacts the individual therapeutic strategy in HCC patients[32]. In recent years, several studies have explored CT or MRI-based radiomics for the prediction of tumor response of HCC after TACE[3337]. Chen et al.[33] developed a CT-based radiomics model on tumoral region for predicting TACE response, and their model merely showed moderate predictive performance with the AUCs of 0.760 and 0.720 in the training and testing cohorts, respectively. MRI-based radiomics might be more promising for response estimation due to better soft-tissue contrast and resolution in MR images. According to the related studies of Sun[34], Kuang[35], Kong[36], and Liu[37], they constructed radiomics models based on preoperative single MRI sequence or multiparametric MRI (MP-MRI) to predict tumor response of HCC patients receiving TACE, and the AUCs ranged from 0.692 to 0.866 in the validation cohort; however, their studies only focused on the intratumoral radiomics features of the HCC. Pathologically, peritumoral parenchyma is representative of cancerous heterogeneity, and the crucial information can be indicated by changes in the area surrounding tumors, such as biological aggressiveness, microinvasion, and micrometastasis[20, 23, 38]; thus, accurate evaluation of the neighboring tissue around tumors may also be useful in predicting treatment response and prognosis of TACE in patients with HCC.

Previous studies have reported that the regions on intratumoral and 3 mm peritumoral, intratumoral and 5 mm peritumoral, and intratumoral and 10 mm peritumoral can provide valuable information for prognosis prediction in HCC[23, 24, 39]. In addition, according to the practice guidelines for the pathological diagnosis of primary liver cancer (2015 update)[40], the liver tissues within 10 mm surrounding the tumor are defined as the adjacent areas around the cancer, where the probability of MVI is high. Therefore, our study focused on predicting treatment response, and we selected the most stable and predictive radiomics features from intratumoral, 3 mm, 5 mm and 10 mm peritumoral regions based on preoperative CE-MR images for radiomics models construction, which can quantitatively assess the heterogeneity and invasiveness of intratumoral and peritumoral tissues in a non-invasive way. In the present study, the PTR (3 mm) and PTR (5 mm) radiomics models showed comparable performance compared with the TR model, which indicated that peritumoral tissues were possess of a clinical value in assessing treatment response. Hu et al.[41] reached a similar result that the intratumoral radiomics model based on enhanced CT images obtained the equivalent efficacy compared with peritumoral radiomics model in predicting the response to neoadjuvant radiotherapy and chemotherapy in esophageal cancer with the AUCs of 0.730 and 0.734, respectively. The radiomics features contributed to the peritumoral models construction in the study were most derived from the AP images. This finding was in agreement with previous studies, in which the presence of peritumoral enhancement in AP images indicated more aggressive biological behavior[23, 42]. Additionally, the PTR (10 mm) radiomics model showed lower predictive performance than the PTR (3 mm) and PTR (5 mm) models. This may be due to the decrease in the number of microsatellites in the farther peritumoral area and inclusion of the peritumoral normal hepatic parenchyma[39]. We further combined intratumoral and peritumoral rad-scores to establish T-PTR (3 mm), T-PTR (5 mm), and T-PTR (10 mm) radiomics models for predicting TACE response, and the T-PTR (3 mm) radiomics model achieved the best-performing performance among the seven radiomics models mentioned above. A similar study reported by Kim et al.[39] found that the radiomics model with intratumoral and 3 mm border extension showed improved performance than the radiomics model with intratumoral and 5 mm border extension for predicting ER of HCC after curative resection. The T-PTR (3 mm) radiomics model demonstrated better predictive performance compared with the TR radiomics model, which indicated that peritumoral radiomics might potentially enhance the ability of intratumoral radiomics for TACE response prediction. This might be interpreted that arterial peritumoral enhancement and irregular margin presented in the peritumoral area are independent predictors of prognosis in HCC patients[23, 39]. Only one of the published studies, conducted MRI-based radiomics on intratumoral and peritumoral regions for the prediction of TACE prognosis[23]. In their study, the radiomics models based on the entire tumor volumetric of AP (APETV), PVPETV, and the border extensions of 1 mm, 3 mm, and 5 mm on the PVP (PVPB1, PVPB3, and PVPB5) were constructed to predict RFS of HCC patients after TACE. The best C-index results for PVPETV and PVPB3 radiomics models were 0.727 and 0.714 in the validation dataset, respectively. However, the above research analyzed the image information of whole areas including intratumoral and peritumoral regions, and did not explore the individual contribution of the area around the tumor to the prediction model; thus, it was unable to determine the significance of the separate peritumoral region in predicting recurrence or prognosis. Compared with the previous study[23], our study may have the following advantages: first, radiomics features derived from three-phase enhanced MR images might more fully reflect tumor heterogeneity and vascularization patterns, which is helpful for efficacy estimation; second, the individual peritumoral (3 mm, 5 mm, and 10 mm) radiomics models were constructed, and the proper and valuable peritumoral distance was determined; third, intratumoral combined peritumoral radiomics analysis may contain more prognostic information, and potentially provide a more accurate and effective approach of individualized efficacy prediction for HCC patients.

In this study, during the construction of the clinical-radiological model, ALP value, tumor size, and satellite nodule were independent predictors associated with treatment response of HCC after TACE. Previous researches on TACE clarified that a higher ALP value was an independent risk factor for unfavourable OS[43, 44]. Our study showed that abnormal ALP value was a significant predictor for poor response of HCC. Serum ALP levels are usually elevated in patients with liver diseases and thus may reflect the status of liver injury. In addition, ALP has already been included in the Chinese University Prognostic Index, a HCC staging system that assigns a score of 3 when ALP is > 200 IU/L, indicating the potential role of ALP in predicting the prognosis of HCC patients[45]. Tumor size has been broadly recognized as a major predictive factor of treatment response for TACE[10, 32]. Larger tumors usually have more satellite lesions or daughter nodules making it difficult for TACE to achieve CR[46]. In our study, maximal tumor size > 5 cm was a significant predictive factor for NR, a result similar to the study by Jeong et al.[47], who reported that tumors > 5 cm were independently associated with failure to achieve CR after TACE. Several studies reported that satellite nodule surrounding the main tumor was closely related to tumor grade, MVI, and ER after curative resection and OS of TACE therapy[22, 37, 48]. Our study demonstrated that the presence of satellite nodule was inclined to show NR to TACE, which was consistent with a recent study reported by Li et al.[11], their study showed that satellite nodule was statistically significant between the OR and NR groups. This may be interpreted that the development of satellite nodule favors vascular invasion and also tumor recurrence[48].

We ultimately developed a combined nomogram by integrating the T-PTR (3 mm) rad-score with clinical-radiological risk indicators (ALP value, tumor size, and satellite nodule) for treatment response prediction. The combined nomogram achieved good calibration and the strongest predictive performance based on AUCs in the training (nomogram vs. radiomics model vs. clinical-radiological model, 0.910 vs. 0.884 vs. 0.789) and validation (nomogram vs. radiomics model vs. clinical-radiological model, 0.918 vs. 0.911 vs. 0.782) cohorts. The novel combined prediction nomogram was evaluated by a decision curve to clarify the clinical usefulness, which could offer insight into clinical outcomes on the basis of threshold probability, from which the net benefit could be derived[36]. Our results clearly demonstrated that the combined nomogram could obtain more net benefit than either the treat-all-patients or the treat-none-patients strategies across a wide range of threshold probabilities. Therefore, our novel nomogram may provide a reliable and efficient tool that enables visualized and personalized decision-making for the treatment management of patients with HCC.

This study has several limitations. First, this was a retrospective study with a limited patient cohort at a single center, which may introduce selection bias. A larger database from multiple centers is further needed to externally validate the robustness and reproducibility of the prediction models. Second, ROIs were delineated manually by radiologists, and thus is time-consuming and prone to error and user variability; thus, it’s essential to develop an automatic and reliable liver tumor segmentation tool. Third, it should be noted that the MP-MRI data are not included in this study. In the future, we will attempt to develop a radiomics approach based on MP-MRI for response evaluation in HCC patients after TACE. Finally, we did not include genomics hallmarks related with TACE response, which may provide additional valuable information for therapeutic efficacy prediction. A future direction to consider would be identification of genetic features of HCCs that can be incorporated into the prediction schema.

Conclusions

In conclusion, intratumoral and peritumoral radiomics based on preoperative CE-MR images can enhance the ability of radiomics model in predicting tumor response to TACE. The combined nomogram which incorporated the rad-score and clinical-radiological risk factors provides an effective tool for the precise and individualized estimation of treatment response for HCC patients treated with TACE. The accurate identification of HCC patients who would receive benefit from upfront TACE might potentially help decision-making for subsequent treatment strategies.

Abbreviations

AASLD, American Association for the Study of Liver Disease; AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALB, albumin; ALP, alkaline phosphatase; AP, arterial phase; BCLC, Barcelona Clinic Liver Cancer; CR, complete response; CE-MRI, contrast-enhanced magnetic resonance imaging; CI, confidence interval; DP, delayed phase; DCA, decision curve analysis; ECOG, Eastern Cooperative Oncology Group; ER, early recurrence; GLCM, grey level co-occurrence matrix; GLRLM, grey level run length matrix; GLZSM, grey-level zone size matrix; GGT, γ-glutamyltranspeptadase; GBDT, gradient boosting decision tree; HCC, hepatocellular carcinoma; ICC, intraclass correlation coefficient; mRECIST, modified Response Evaluation Criteria in Solid Tumors; MVI, microvascular invasion; MP-MRI, multiparametric magnetic resonance imaging; NR, non-response; OR, objective response; OS, overall survival; PR, partial response; PD, progression disease; PVP, portal venous phase; PLT, platelet count; PT, prothrombin time; PTR, peritumoral region; Rad-score, radiomics score; RFS, recurrence-free survival; SD, stable disease; TACE, transarterial chemoembolization; TBIL, total bilirubin; T-PTR, intratumoral and peritumoral region; TR, intratumoral region.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Institutional Review Board of the First Affiliated Hospital of Dalian Medical University, and was carried out in accordance with the Declaration of Helsinki. The requirement for the informed consent was waived.

Consent for publication

Not applicable.

Availability of data and materials

All data generated or analysed during this study are included in supplementary material of this article.

Competing interests

None.

Funding

The study was supported by the National Natural Science Foundation of China (No. 61971091).

Authors' contributions

Conceptualization, Y.Z., J.Z., and A.L.L.; Methodology, Y.Z. and Y.G.; Software, Y.G.; Validation, Y.Z., N.W., and Q.H.X.; Formal analysis, Y.Z., N.W., Y.H.L., and J.H.L.; Investigation, Y.Z., J.Z., N.W., Q.H.X., Y.H.L., J.H.L., X.Y.Z., A.L.C., and L.H.C; Resources, Y.Z., J.Z., and N.W.; Data curation, Y.Z., N.W., Q.H.X., J.H.L., and Q.H.Z.; Writing-original draft preparation, Y.Z. and J.Z.; Writing-review and editing, F.W., Y.G., and A.L.L.; Visualization, Y.Z., L.J.S, and Y.G.; Supervision, A.L.L.; Project administration, Q.W.S. and A.L.L.; Funding acquisition, A.L.L.. All authors have read and agreed to the published version of the manuscript.

Acknowledgements

Not applicable. 

Authors' information (optional)

1Department of Radiology, The First Affiliated Hospital, Dalian Medical University, Dalian, China; 2Department of Interventional Radiology, The First Affiliated Hospital, Dalian Medical University, Dalian, China; 3College of Medical Imaging, Dalian Medical University, Dalian, China; 4GE Healthcare (China), Shanghai, China.

 

References

  1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71:209-249.
  2. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin. 2019;69:7-34.
  3. Lei Z, Li J, Wu D, Xia Y, Wang Q, Si A, et al. Nomogram for preoperative estimation of microvascular invasion risk in hepatitis B virus-related hepatocellular carcinoma within the milan criteria. JAMA Surg. 2016;151:356-363.
  4. Lin S, Hofmann K, Schemmer P. Treatment of hepatocellular carcinoma: a systematic review. Liver Cancer. 2012;1:144-158.
  5. Chang Y, Jeong SW, Young Jang J, Jae Kim Y. Recent Updates of Transarterial Chemoembolilzation in Hepatocellular Carcinoma. Int J Mol Sci. 2020;21:8165.
  6. Piscaglia F, Ogasawara S. Patient Selection for Transarterial Chemoembolization in Hepatocellular Carcinoma: Importance of Benefit/Risk Assessment. Liver Cancer. 2018;7:104-119.
  7. Guan YS, He Q, Wang MQ. Transcatheter arterial chemoembolization: History for more than 30 years. ISRN Gastroenterol. 2012;2012:480650.
  8. Song YG, Shin SW, Cho SK, Choi D, Rhim H, Lee MW, et al. Transarterial chemoembolization as first-line therapy for hepatocellular carcinomas infeasible for ultrasound-guided radiofrequency ablation: a retrospective cohort study of 116 patients. Acta Radiol. 2015;56:70-77.
  9. Zhang H, He X, Yu J, Song W, Liu X, Liu Y, et al. Preoperative MRI features and clinical laboratory indicators for predicting the early therapeutic response of hepatocellular carcinoma to transcatheter arterial chemoembolization combined with High-intensity focused ultrasound treatment. Br J Radiol. 2019;92:20190073.
  10. Kim YJ, Lee MH, Choi SY, Yi BH, Lee HK. Magnetic resonance imaging features predictive of an incomplete response to transarterial chemoembolization in patients with hepatocellular carcinoma: A STROBE-compliant study. Medicine (Baltimore). 2019;98:e15592.
  11. Li ZW, Ren AH, Yang DW, Xu H, Wei J, Yuan CW, et al. Preoperatively predicting early response of HCC to TACE using clinical indicators and MRI features. BMC Med Imaging. 2022;22:176.
  12. Wu XM, Wang JF, Ji JS, Chen MG, Song JG. Evaluation of efficacy of transcatheter arterial chemoembolization for hepatocellular carcinoma using magnetic resonance diffusion-weighted imaging. Onco Targets Ther. 2017;10:1637-1643.
  13. Wu L, Xu P, Rao S, Yang L, Chen C, Liu H, et al. ADCtotal ratio and D ratio derived from intravoxel incoherent motion early after TACE are independent predictors for survival in hepatocellular carcinoma. J Magn Reson Imaging. 2017;46:820-830.
  14. Yuan ZG, Wang ZY, Xia MY, Li FZ, Li Y, Shen Z, et al. Diffusion Kurtosis Imaging for Assessing the Therapeutic Response of Transcatheter Arterial Chemoembolization in Hepatocellular Carcinoma. J Cancer. 2020;11:2339-2347.
  15. Saito K, Ledsam J, Sugimoto K, Sourbron S, Araki Y, Tokuuye K. DCE-MRI for Early Prediction of Response in Hepatocellular Carcinoma after TACE and Sorafenib Therapy: A Pilot Study. J Belg Soc Radiol. 2018;102:40.
  16. Aerts HJ. The potential of radiomic-based phenotyping in precision medicine: a review. JAMA Oncol. 2016;2:1636-1642.
  17. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14:749-762.
  18. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun. 2014;5:4006.
  19. Zhao Y, Chen R, Zhang T, Chen C, Muhelisa M, Huang J, et al. MRI-Based Machine Learning in Differentiation Between Benign and Malignant Breast Lesions. Front Oncol. 2021;11:552634.
  20. Tian Y, Hua H, Peng Q, Zhang Z, Wang X, Han J, et al. Preoperative Evaluation of Gd-EOB-DTPA-Enhanced MRI Radiomics-Based Nomogram in Small Solitary Hepatocellular Carcinoma (≤3 cm) With Microvascular Invasion: A Two-Center Study. J Magn Reson Imaging. 2022;56:1459-1472.
  21. Shin J, Seo N, Baek SE, Son NH, Lim JS, Kim NK, et al. MRI Radiomics Model Predicts Pathologic Complete Response of Rectal Cancer Following Chemoradiotherapy. Radiology. 2022;303:351-358.
  22. Ji GW, Zhu FP, Xu Q, Wang K, Wu MY, Tang WW, et al. Radiomic Features at Contrast-enhanced CT Predict Recurrence in Early Stage Hepatocellular Carcinoma: A Multi-Institutional Study. Radiology. 2020;294:568-579.
  23. Song W, Yu X, Guo D, Liu H, Tang Z, Liu X, et al. MRI-Based Radiomics: Associations With the Recurrence-Free Survival of Patients With Hepatocellular Carcinoma Treated With Conventional Transcatheter Arterial Chemoembolization. J Magn Reson Imaging. 2020;52:461-473.
  24. Chong H, Gong Y, Pan X, Liu A, Chen L, Yang C, et al. Peritumoral Dilation Radiomics of Gadoxetate Disodium-Enhanced MRI Excellently Predicts Early Recurrence of Hepatocellular Carcinoma without Macrovascular Invasion After Hepatectomy. J Hepatocell Carcinoma. 2021;8:545-563.
  25. Zhu XD, Zhang JB, Zhuang PY, Zhu HG, Zhang W, Xiong YQ, et al. High expression of macrophage colony-stimulating factor in peritumoral liver tissue is associated with poor survival after curative resection of hepatocellular carcinoma. J Clin Oncol. 2008;26:2707-2716.
  26. Cheng Z, Yang P, Qu S, Zhou J, Yang J, Yang X, et al. Risk factors and management for early and late intrahepatic recurrence of solitary hepatocellular carcinoma after curative resection. HPB (Oxford). 2015;17:422-427.
  27. Dai XM, Huang T, Yang SL, Zheng XM, Chen GG, Zhang T. Peritumoral EpCAM Is an Independent Prognostic Marker after Curative Resection of HBV-Related Hepatocellular Carcinoma. Dis Markers. 2017;2017:8495326.
  28. Dai X, Xue J, Hu J, Yang SL, Chen GG, Lai PBS, et al. Positive Expression of Programmed Death Ligand 1 in Peritumoral Liver Tissue is Associated with Poor Survival after Curative Resection of Hepatocellular Carcinoma. Transl Oncol. 2017;10:511-517.
  29. Kong LQ, Zhu XD, Xu HX, Zhang JB, Lu L, Wang WQ, et al. The Clinical Significance of the CD163+ and CD68+ Macrophages in Patients with Hepatocellular Carcinoma. PLoS One. 2013;8:e59771.
  30. Marrero JA, Kulik LM, Sirlin CB, Zhu AX, Finn RS, Abecassis MM, et al. Diagnosis, staging and management of hepatocellular carcinoma: 2018 Practice Guidance by the American Association for the Study of Liver Diseases. Hepatology. 2018;68:723-750.
  31. Lencioni R, Llovet JM. Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin Liver Dis. 2010;30:52-60.
  32. Kim BK, Kim SU, Kim KA, Chung YE, Kim MJ, Park MS, et al. Complete response at first chemoembolization is still the most robust predictor for favorable outcome in hepatocellular carcinoma. J Hepatol. 2015;62:1304-1310.
  33. Chen M, Cao J, Hu J, Topatana W, Li S, Juengpanich S, et al. Clinical-Radiomic Analysis for Pretreatment Prediction of Objective Response to First Transarterial Chemoembolization in Hepatocellular Carcinoma. Liver Cancer. 2021;10:38-51.
  34. Sun Y, Bai H, Xia W, Wang D, Zhou B, Zhao X, et al. Predicting the Outcome of Transcatheter Arterial Embolization Therapy for Unresectable Hepatocellular Carcinoma Based on Radiomics of Preoperative Multiparameter MRI. J Magn Reson Imaging. 2020;52:1083-1090.
  35. Kuang Y, Li R, Jia P, Ye W, Zhou R, Zhu R, et al. MRI-Based Radiomics: Nomograms predicting the short-term response after transcatheter arterial chemoembolization (TACE) in hepatocellular carcinoma patients with diameter less than 5 cm. Abdom Radiol (NY). 2021;46:3772-3789.
  36. Kong C, Zhao Z, Chen W, Lv X, Shu G, Ye M, et al. Prediction of tumor response via a pretreatment MRI radiomics-based nomogram in HCC treated with TACE. Eur Radiol. 2021;31:7500-7511.
  37. Liu QP, Yang KL, Xu X, Liu XS, Qu JR, Zhang YD. Radiomics analysis of pretreatment MRI in predicting tumor response and outcome in hepatocellular carcinoma with transarterial chemoembolization: a two-center collaborative study. Abdom Radiol (NY). 2022;47:651-663.
  38. Budhu A, Forgues M, Ye QH, Jia HL, He P, Zanetti KA, et al. Prediction of venous metastases, recurrence, and prognosis in hepatocellular carcinoma based on a unique immune response signature of the liver microenvironment. Cancer Cell. 2006;10:99-111.
  39. Kim S, Shin J, Kim DY, Choi GH, Kim MJ, Choi JY. Radiomics on Gadoxetic Acid-Enhanced Magnetic Resonance Imaging for Prediction of Postoperative Early and Late Recurrence of Single Hepatocellular Carcinoma. Clin Cancer Res. 2019;25:3847-3855.
  40. Cong WM, Bu H, Chen J, Dong H, Zhu YY, Feng LH, et al; Guideline Committee. Practice guidelines for the pathological diagnosis of primary liver cancer: 2015 update. World J Gastroenterol. 2016;22:9279-9287.
  41. Hu Y, Xie C, Yang H, Ho JWK, Wen J, Han L, et al. Assessment of Intratumoral and Peritumoral Computed Tomography Radiomics for Predicting Pathological Complete Response to Neoadjuvant Chemoradiation in Patients With Esophageal Squamous Cell Carcinoma. JAMA Netw Open. 2020;3:e2015927.
  42. Meng XP, Wang YC, Ju S, Lu CQ, Zhong BY, Ni CF, et al. Radiomics Analysis on Multiphase Contrast-Enhanced CT: A Survival Prediction Tool in Patients With Hepatocellular Carcinoma Undergoing Transarterial Chemoembolization. Front Oncol. 2020;10:1196.
  43. Le Y, Shen JX, Zhang YF, He MK, Kan A, Chen HL, et al. Transarterial Chemoembolization related to Good Survival for Selected Patients with advanced Hepatocellular Carcinoma. J Cancer. 2019;10:665-671.
  44. Wang H, Du PC, Wu MC, Cong WM. Postoperative adjuvant transarterial chemoembolization for multinodular hepatocellular carcinoma within the Barcelona Clinic Liver Cancer early stage and microvascular invasion. Hepatobiliary Surg Nutr. 2018;7:418-428.
  45. Leung TW, Tang AM, Zee B, Lau WY, Lai PB, Leung KL, et al. Construction of the Chinese University Prognostic Index for hepatocellular carcinoma and comparison with the TNM staging system, the Okuda staging system, and the Cancer of the Liver Italian Program staging system: a study based on 926 patients. Cancer. 2002;94:1760-1769.
  46. Bannangkoon K, Hongsakul K, Tubtawee T, McNeil E, Sriplung H, Chongsuvivatwong V. Rate and Predictive Factors for Sustained Complete Response after Selective Transarterial Chemoembolization (TACE) in Patients with Hepatocellular Carcinoma. Asian Pac J Cancer Prev. 2018;19:3545-3550.
  47. Jeong SO, Kim EB, Jeong SW, Jang JY, Lee SH, Kim SG, et al. Predictive Factors for Complete Response and Recurrence after Transarterial Chemoembolization in Hepatocellular Carcinoma. Gut Liver. 2017;11:409-416.
  48. Ahn SJ, Kim JH, Park SJ, Kim ST, Han JK. Hepatocellular carcinoma: preoperative gadoxetic acid-enhanced MR imaging can predict early recurrence after curative resection using image features and texture analysis. Abdom Radiol (NY). 2019;44:539-548.