Abstract
Objective
The optimal assessment of tumor response in hepatocellular carcinoma (HCC) after drug-eluting bead transarterial chemoembolization (DEB-TACE) remains unclear. This study aimed to develop a CT-based radiomics model for the preoperative prediction of tumor response to DEB-TACE in patients with HCC.
Methods
Patients with HCC who received DEB-TACE as initial treatment from two centers were included and divided into training, internal validation and external validation cohorts. LASSO and logistic regression were used to identify the optimal radiomics features and independent predictors of tumor response, respectively, Then, a combination model was developed and assessed.
Results
335 patients were included in this study. Radscore was calculated based on 15 identified optimal radiomics features, and maximum tumor diameter and tumor capsule were independent predictors of tumor response. A nomogram was generated based on the clinical predictors and Radscore. In the training cohort, the AUC of nomogram was significantly superior to that of both clinical (0.915 vs 0.800, P=0.004) and the radiomics models (0.915 vs 0.842, P=0.010). Calibration curves and decision curve analysis (DCA) demonstrated good consistency between the nomogram predictions and actual outcomes as well as clinical net benefit across all cohorts. The prognostic stratification based on the nomogram effectively predicted patients with potential objective response of tumor and survival.
Conclusion
This radiomics model had an excellent performance to predict tumor response in HCC after DEB-TACE, which may serve as a reliable tool to assist with the selection of patients for DEB-TACE.
Keywords: radiomics, prediction, hepatocellular carcinoma, transarterial chemoembolization
Introduction
Hepatocellular carcinoma (HCC) ranks among the most prevalent malignancies globally and is a leading cause of cancer-related mortality.1,2 According to the Barcelona Clinic Liver Cancer (BCLC) staging system, transarterial chemoembolization (TACE) is recommended as the mainstream treatment for intermediate-stage HCC.3 Compared with conventional TACE (cTACE) using Lipiodol as embolic agent, drug-eluting bead TACE (DEB-TACE) exhibits enhanced tolerability and reduced incidence of side effects, primarily due to its capacity for slow drug release, thereby reducing systemic drug concentration and minimizing systemic adverse reactions. Furthermore, DEB-TACE demonstrates significant advantages in improving tumor response and progression-free survival (PFS).4 Therefore, DEB-TACE serves as a more desirable option for the treatment of unresectable HCC.
However, the differential efficacy of DEB-TACE was observed due to the different physical status, liver function, and tumor burden of individual patient.5 Therefore, it is crucial for clinicians to identify patients who are likely to derive the most significant survival benefit from DEB-TACE. Currently, various staging and scoring systems are utilized to objectively assess the condition of patients with HCC for outcomes prediction, such as BCLC, China liver cancer staging (CNLC) system5 and hepatoma arterial-embolization prognostic (HAP) score.6 Regrettably, these systems are not specifically designed for HCC undergoing DEB-TACE without the radiologic characteristics of tumor, which may lead to a limitation in prognostic evaluation and in identifying subgroups of patients who might benefit from DEB-TACE.
Preoperative radiomics model is gaining increasing attention in the tumor response and prognostic prediction of patients, which extracts quantitative features from image data.7,8 Radiomic features extracted from CT images are inextricably linked to intratumoural and intertumoural heterogeneity and provide important information for histologic grading of HCC, microvascular invasion status.9 Therefore, it may be feasible to predict the tumor response after DEB-TACE by using radiomics.
This study aimed to develop a preoperative model based on CT imaging features for prediction of the objective response of HCC to DEB-TACE. The role of model was externally validated in identifying patients with HCC likely to benefit from DEB-TACE as their initial treatment, and in guiding the formulation of personalized treatment and follow-up strategy.
Materials and Methods
Patients
We included patients diagnosed with HCC between June 2016 and December 2020 from two medical centers. Clinical data were retrospectively collected from 268 primary HCC patients who underwent DEB-TACE at one center. These patients were randomly allocated to a training cohort and an internal validation cohort in 7:3 ratio. Additional patients from another center served as an external validation cohort. Diagnosis for HCC was confirmed by a positive finding on two imaging modalities or one imaging modality combined with an alpha-fetoprotein (AFP) level exceeding 400 ng/mL, or cytological/histological confirmation.
The inclusion criteria included: (1) age between 18 and 80 years; (2) a diagnosis of unresectable HCC; (3) Eastern Cooperative Oncology Group (ECOG) performance status scores ranging from 0 to 2; (4) Child-Pugh class of A or B; (5) liver tumor had ≥1 measurable lesion with a long diameter ≥10 mm for the measurable lesion, as well as accounting for <60% of total liver volume; and (6) availability of abdominal contrast-enhanced CT within 1 week prior to DEB-TACE for subsequent radiomics analysis. The exclusion criteria included: (1) any other anti-tumor therapy before DEB-TACE; (2) presence of distant metastasis; and (3) incomplete records.
Dynamic enhanced CT or MRI was conducted 4 weeks after DEB-TACE and every 2–3 months thereafter. Tumor response to DEB-TACE was evaluated based on the difference between pre-operative and post-operative contrast-enhanced CT/MRI using the modified Response Evaluation Criteria in Solid Tumors (mRECIST). In patients undergoing multiple DEB-TACE procedures, the best response achieved during the course of treatment was used to estimate the tumor response to DEB-TACE. Complete response (CR) was defined as disappearance of any intratumoral arterial enhancement in all target lesions; partial response (PR) as at least a 30% decrease in the sum of diameters of viable (enhancing) target lesions compared with baseline; progressive disease (PD) as at least a 20% increase in the sum of diameters of viable target lesions or the appearance of new lesions; and stable disease (SD) as a change that falls between the thresholds for PR and PD. Baseline clinical characteristics, including age, sex, ECOG performance status score, hepatitis virus status, AFP level, liver function, presence of liver cirrhosis and Child-Pugh class, were extracted from medical records.
This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University (2024-KY-1588-002) in compliance with the Declaration of Helsinki. The requirement for informed consent to participate was waived by Ethics Committee due to its retrospective nature. All clinical data used in this study were anonymous.
DEB-TACE Procedure
All patients at two centers were administered standardized DEB-TACE. Using a transfemoral approach, a 5-F RH catheter (Cook) was inserted into the celiac trunk or variable vessels for angiography to define the tumor-feeding arteries. Superselective catheterization was performed with a 2.7-F Progreat microcatheter (Terumo) into the distal tumor-feeding arteries. Via the microcatheter, 100–300 or 300–500 μm CalliSpheres microspheres (Jiangsu Hengrui Kellison Biomedical Co., Ltd., Suzhou, China) were injected loaded with 60 mg of pirarubicin or epirubicin. If the tumor staining was still present at repeated imaging after DEBs, 350–560 μm or 560–710 μm gelatin sponge particles or polyvinyl alcohol (ALICON Pharm SCI & TECH) were added until the disappearance of tumor staining, which was defined as complete absence of intratumoral contrast staining on the final angiographic run compared with the pre-embolization angiogram. These procedures were performed by two interventional radiologists with more than 10 years of experience.
CT Image Acquisition and Radiologic Evaluation
Procedures for CT image acquisition are described in detail in Supplementary material. Two radiologists with 5 (Radiologist A) and 10 years (Radiologist B) of experience in abdominal CT imaging respectively independently reviewed all images, blinded to clinical and follow-up information. They evaluated maximum tumor diameter, number of lesions, presence of tumor capsule, portal vein invasion, and arteriovenous fistula. Quantitative measurements were averaged between the two radiologists. In cases of qualitative discrepancies, a third senior radiologist with over 20 years of experience was consulted for consensus.
Image Segmentation and Feature Extraction
Image segmentation was performed on arterial phase CT images, which were selected for their superior contrast between hypervascular tumors and the liver parenchyma, ensuring greater robustness for segmentation. Using the open-source 3D Slicer software (version 5.2.1; https://www.slicer.org/), the volume of interest (VOI) for each patient was delineated via semi-automatic segmentation. This process, guided by mRECIST criteria, ensured the VOI collectively encompassed the entire tumor burden scheduled for TACE, including all lesions in cases of multifocal disease.
To mitigate the impact of varying scanning protocols, the delineated VOIs were resampled to a 1×1×1 mm3 voxel size, and gray levels were discretized using a bin width of 25. Following this preprocessing, a total of 851 radiomics features were extracted per patient, comprising 18 first-order, 14 shape, and 75 original textural features, as well as 744 corresponding wavelet features. The 75 textural features included 24 from the gray-level co-occurrence matrix (GLCM), 14 from the gray-level dependence matrix (GLDM), 16 from the gray-level run-length matrix (GLRLM), 16 from the gray-level size zone matrix (GLSZM), and 5 from the neighborhood gray-tone difference matrix (NGTDM). Finally, all extracted features were standardized via z-score normalization. This entire workflow for image processing and feature extraction adhered strictly to the Image Biomarker Standardization Initiative (IBSI) guidelines to ensure robustness and reproducibility.
Subsequently, radiomics features were extracted from the resulting VOI for each patient, facilitating a per-patient basis for analysis.
To ensure segmentation consistency, all primary VOIs for all cohorts were delineated by Radiologist A. The robustness of this process was subsequently assessed on a randomly selected subset of 30 patients to determine intra- and inter-observer reliability. For intra-observer analysis, Radiologist A repeated the segmentations after a two-week interval. For inter-observer analysis, Radiologist B independently delineated the VOIs for the same cohort.
Feature Selection, Model Construction and Evaluation
We employed a three-step feature selection process on the training cohort data: (1) features with inter- and intra-observer intraclass correlation coefficients (ICCs) > 0.75 were retained; (2) highly correlated feature pairs (correlation coefficient ≥ 0.90) were pruned by removing the feature with the higher average correlation; (3) the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm was applied to select the most valuable subset of features. The LASSO regularization parameter (λ) was determined through 10-fold cross-validation, minimizing the average mean square error. The selected features were used to construct a radiomics signature within the training cohort.
For clinical characteristics, univariate logistic regression analysis was performed, with variables having P < 0.10 selected for multivariate logistic regression to identify independent predictors of DEB-TACE response. We developed a clinical model using independent clinical predictors, a radiomics model using the radiomics signature, and a combined model incorporating both. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), with comparisons made using the DeLong test. Based on the combined model, a nomogram was constructed to provide a visual tool for predicting tumor response to DEB-TACE. To further verify the clinical utility of the constructed nomogram and enable comparative analysis with existing classic models, three widely used clinical staging/scoring systems (BCLC, CNLC, and mHAP-II) were included in the analysis. Their predictive efficacy for DEB-TACE response was evaluated using ROC curve analysis, following the same assessment criteria as the aforementioned models.
Additionally, the predictive accuracy of the nomogram was evaluated using calibration curves in the training, internal validation, and external validation cohorts. The Hosmer-Lemeshow test was employed to assess the goodness-of-fit of the nomogram. Additionally, decision curve analysis (DCA) was conducted to assess the clinical net benefit of the nomogram across different threshold probabilities in all three cohorts.
Prognostic Value of the Nomogram
The prognostic utility of the nomogram was evaluated by assessing its association with overall survival (OS). Survival data were obtained through a review of medical records and telephone follow-ups, with the final follow-up conducted on 31/12/2020 for both training and internal validation cohorts, and 30/9/2021 for the external validation cohort. OS was defined as the interval from the date of the initial DEB-TACE procedure to the date of death or the last follow-up. For survival analysis, patients were stratified into two groups based on the nomogram. Kaplan-Meier analysis and the Log rank test were used to compare OS between these groups.
Statistical Analysis
All statistical analyses were performed using R software (version 4.2.2, https://www.r-project.org/). Normally distributed continuous variables were reported as mean ± standard deviation, while non-normally distributed variables were expressed as median (interquartile range). Categorical variables were analyzed using chi-square tests or Fisher’s exact tests. Continuous variables were compared using Student’s t-test or Mann–Whitney U-test, as appropriate. Statistical significance was set at P<0.05.
Results
Patient Characteristics
There were 189, 79 and 67 patients in the training cohort, internal validation cohort and external validation cohort. The patient inclusion process is illustrated in Figure 1. Table 1 summarizes the baseline characteristics of the patients across all cohorts. As for DEB-TACE protocol, 100–300μm of DEBs were used in 52.9% and 76.1% of patients and 300–500μm in 47.0% and 23.8% of patients at two centers, respectively (P<0.001). In addition, pirarubicin was used in 41.0% and 47.8% of patients and epirubicin in 59.0% and 52.2% of patients at two centers, respectively (P>0.05).
Figure 1.
Patient flowchart for this study.
Table 1.
Baseline Characteristics of Patients with Hepatocellular Carcinoma
| Characteristic | Training Cohort (n = 189) | Internal Verification (n = 79) | External Verification (n = 67) | ||||||
|---|---|---|---|---|---|---|---|---|---|
| CR+PR | PD+SD | P value | CR+PR | PD+SD | P value | CR+PR | PD+SD | P value | |
| Age (y)* | 55.74 ± 10.56 | 51.73 ± 12.47 | 0.130 | 55.88 ± 10.22 | 46.6 ± 6.88 | 0.002 | 62.47 ± 10.48 | 53.3 ± 13.2 | |
| Sex | 0.774 | > 0.999 | > 0.999 | ||||||
| Men | 137 (84.05%) | 23 (88.46%) | 59 (85.51%) | 9 (90%) | 43 (91.49%) | 19 (95%) | |||
| Women | 26 (15.95%) | 3 (11.54%) | 10 (14.49%) | 1 (10%) | 4 (8.51%) | 1 (5%) | |||
| PS score | 0.070 | 0.383 | > 0.999 | ||||||
| 0 | 125 (76.69%) | 24 (92.31%) | 58 (84.06%) | 10 (100%) | 43 (91.49%) | 19 (95%) | |||
| 1-2 | 38 (23.31%) | 2 (7.69%) | 11 (15.94%) | 0 (0%) | 4 (8.51%) | 1 (5%) | |||
| Hepatitis virus | 0.970 | 0.383 | > 0.999 | ||||||
| HBV | 142 (87.12%) | 22 (84.62%) | 58 (84.06%) | 10 (100%) | 39 (82.98%) | 17 (85%) | |||
| Others | 21 (12.88%) | 4 (15.38%) | 11 (15.94%) | 0 (0%) | 8 (17.02%) | 3 (15%) | |||
| AFP | 0.570 | 0.640 | 0.513 | ||||||
| Normal | 85 (52.15%) | 12 (46.15%) | 37 (53.62%) | 4 (40%) | 38 (80.85%) | 14 (70%) | |||
| Abnormal | 78 (47.85%) | 14 (53.85%) | 32 (46.38%) | 6 (60%) | 9 (19.15%) | 6 (30%) | |||
| ALT | 0.273 | 0.640 | 0.465 | ||||||
| Normal | 100 (61.35%) | 13 (50%) | 37 (53.62%) | 4 (40%) | 31 (65.96%) | 15 (75%) | |||
| Abnormal | 63 (38.65%) | 13 (50%) | 32 (46.38%) | 6 (60%) | 16 (34.04%) | 5 (25%) | |||
| AST | 0.049 | 0.026 | 0.153 | ||||||
| Normal | 64 (39.26%) | 5 (19.23%) | 29 (42.03%) | 0 (0%) | 30 (63.83%) | 9 (45%) | |||
| Abnormal | 99 (60.74%) | 21 (80.77%) | 40 (57.97%) | 10 (100%) | 17 (36.17%) | 11 (55%) | |||
| Albumin (g/L)* | 37.89 ± 5.01 | 36.94 ± 4.77 | 0.359 | 37.51 ± 6.5 | 37.87 ± 3.74 | 0.804 | 41.3 ± 5.00 | 41.58 ± 5.64 | |
| Total bilirubin (μmol/L)* | 15.54 ± 8.31 | 17.58 ± 11.84 | 0.404 | 16.06 ± 11.21 | 22.4 ± 17.77 | 0.298 | 15.84 ± 7.74 | 20.49 ± 14.17 | |
| Liver Cirrhosis | 0.683 | > 0.999 | 0.294 | ||||||
| Presence | 107 (65.64%) | 16 (61.54%) | 39 (56.52%) | 6 (60%) | 29 (61.7%) | 15 (75%) | |||
| Absence | 56 (34.36%) | 10 (38.46%) | 30 (43.48%) | 4 (40%) | 18 (38.3%) | 5 (25%) | |||
| Child-Pugh | 0.481 | > 0.999 | 0.507 | ||||||
| A | 138 (84.66%) | 20 (76.92%) | 52 (75.36%) | 7 (70%) | 44 (93.62%) | 17 (85%) | |||
| B | 25 (15.34%) | 6 (23.08%) | 17 (24.64%) | 3 (30%) | 3 (6.38%) | 3 (15%) | |||
| Maximum tumor diameter (mm)* | 78.15 ± 37.51 | 115.12 ± 39.11 | < 0.001 | 79.65 ± 35.35 | 108 ± 31.25 | 0.021 | 51.13 ± 26.35 | 58.45 ± 31.48 | |
| Number of lesions* | 1.94 ± 1.36 | 2.27 ± 1.61 | 0.330 | 1.96 ± 1.32 | 4.5 ± 2.68 | 0.015 | 2.04 ± 1.06 | 2 ± 1.59 | |
| Tumor Capsule | 0.005 | 0.383 | 0.165 | ||||||
| Presence | 49 (30.06%) | 1 (3.85%) | 11 (15.94%) | 0 (0%) | 7 (14.89%) | 0 (0%) | |||
| Absence | 114 (69.94%) | 25 (96.15%) | 58 (84.06%) | 10 (100%) | 40 (85.11%) | 20 (100%) | |||
| Portal Vein Invasion | 0.027 | 0.005 | 0.003 | ||||||
| Presence | 57 (34.97%) | 15 (57.69%) | 20 (28.99%) | 8 (80%) | 6 (12.77%) | 10 (50%) | |||
| Absence | 106 (65.03%) | 11 (42.31%) | 49 (71.01%) | 2 (20%) | 41 (87.23%) | 10 (50%) | |||
| Fistula | 0.558 | 0.279 | > 0.999 | ||||||
| Presence | 23 (14.11%) | 2 (7.69%) | 8 (11.59%) | 3 (30%) | 4 (8.51%) | 2 (10%) | |||
| Absence | 140 (85.89%) | 24 (92.31%) | 61 (88.41%) | 7 (70%) | 43 (91.49%) | 18 (90%) | |||
Note: Data are number of patients; data in parentheses are percentage unless otherwise indicated. * Data are means ± standard deviations.
Abbreviations: CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease; AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate transaminase.
In the training cohort, 51 (26.98%), 112 (59.26%), 12 (6.35%) and 14 (7.41%) patients achieved CR, PR, PD and SD according to mRECIST criteria respectively, with 163 patients (86.2%) as responders (CR+PR) and 26 patients (13.8%) as non-responders (SD+PD). The internal validation cohort comprised 69 responders (87.3%, 18 patients CR and 51 patients PR) and 10 non-responders (12.7%, 5 patients SD and 5 patients PD), while the external validation cohort included 47 responders (70.1%, 11 patients CR and 36 patients PR) and 20 non-responders (29.9%, 12 patients SD and 8 patients PD). Significant differences between responders and non-responders were observed in the training cohort for AST levels, maximum tumor diameter, tumor capsule presence, and portal vein invasion. The internal validation cohort showed significant differences in age, AST levels, maximum tumor diameter, number of lesions, and portal vein invasion. In the external validation cohort, only portal vein invasion demonstrated a statistically significant difference between groups.
Feature Selection and Signature Construction
From the initial 851 radiomics features extracted from arterial phase CT images, 456 stable features were retained after ICC and correlation analyses. The LASSO regression algorithm identified 15 optimal radiomics features for constructing the radiomics signature, comprising 1 first-order feature, 3 texture features, and 11 wavelet features (Table 2 and Figure S1). The definitions of 15 selected radiomics features are presented in Supplementary material. The Radscore was calculated for each patient as a weighted sum of the selected features. The formula is presented as follows:
 |
Table 2.
Coefficients of Each Feature in the Radiomics Signature
| Features | Coefficients |
|---|---|
| Intercept | −1.941 |
| original_firstorder_Mean | 0.103 |
| original_gldm_DependenceNonUniformityNormalized | 0.111 |
| original_glcm_ClusterTendency | 3.316 |
| original_glcm_Imc2 | 0.115 |
| wavelet-LLH_glszm_LargeAreaHighGrayLevelEmphasis | −0.023 |
| wavelet-LHH_firstorder_Skewness | 0.086 |
| wavelet-LHH_glszm_SmallAreaLowGrayLevelEmphasis | 0.093 |
| wavelet-LHH_gldm_LargeDependenceEmphasis | 0.221 |
| wavelet-LHH_glcm_Id | 0.074 |
| wavelet-HLL_glszm_LargeAreaLowGrayLevelEmphasis | 0.169 |
| wavelet-HLL_glcm_MCC | −0.319 |
| wavelet-HHL_glcm_ClusterShade | 1.990 |
| wavelet-HHH_glszm_GrayLevelNonUniformity | −0.116 |
| wavelet-HHH_glszm_ZoneEntropy | −0.023 |
| wavelet-HHH_gldm_DependenceNonUniformityNormalized | 0.128 |
The resultant Radscore was significantly lower in the response group compared to the non-response group across all cohorts (P<0.001, Figure 2).
Figure 2.
Distribution of Radscore and inter-group comparisons in three cohorts. (A) Training cohort; (B) Internal validation cohort; (C) External validation cohort.
Models Construction and Evaluation
Logistic regression analysis identified maximum tumor diameter (OR=0.98, P=0.001) and tumor capsule presence (OR=0.12, P=0.046) as independent predictors of tumor response following DEB-TACE (Table 3). These predictors formed the basis of the clinical model. The radiomics model was developed using the Radscore, while the combined model incorporated both the clinical independent predictors and the Radscore. A web nomogram was generated based on the combined model (https://zpc0v0.shinyapps.io/HCC_D-TACE/) (Figure 3). The predictive performance of each model for DEB-TACE treatment response across all cohorts is presented in Table 4, including the three classic staging/scoring systems (BCLC, CNLC, and mHAP-II) added for comparative analysis. In the training cohort, DeLong test results showed that the AUC of the nomogram was significantly superior to both the clinical model (0.915 vs 0.800, P=0.004) and the radiomics model (0.915 vs 0.842, P=0.010) and also significantly superior to the BCLC, CNLC, and mHAP-II models (all P<0.001). These differences were nevertheless attenuated in the validation cohorts. While the DeLong test revealed no statistically significant differences between the nomogram and other models across both validation cohorts, the nomogram maintained superior AUC and other performance metrics relative to all comparators (Figure 4 and Table 5).
Table 3.
Results of Logistic Regression Analysis for Clinical Characteristics
| Characteristic (Reference Level) | Univariate Analysis | Multivariable Analysis | ||||
|---|---|---|---|---|---|---|
| Odds Ratio | 95% CI | P value | Odds Ratio | 95% CI | P value | |
| Age | 1.04 | 0.04–30.91 | 0.083 | 1.01 | 0.47–2.16 | 0.697 |
| Sex (Man) | 1.46 | 0.47–4.51 | 0.564 | |||
| PS score (0) | 3.65 | 0.13–103.20 | 0.088 | 3.13 | 0.24–40.38 | 0.192 |
| Hepatitis virus (Others) | 1.23 | 0.62–2.44 | 0.727 | |||
| AFP (Normal) | 0.79 | 0.26–2.39 | 0.571 | |||
| ALT (Normal) | 0.63 | 0.07–5.33 | 0.276 | |||
| AST (Normal) | 0.37 | 0.01–15.58 | 0.056 | 0.50 | 0.05–5.42 | 0.223 |
| Albumin | 1.04 | 0.18–6.04 | 0.369 | |||
| Total bilirubin | 0.98 | 0.12–8.11 | 0.28 | |||
| Liver Cirrhosis (Absence) | 1.19 | 0.54–2.65 | 0.684 | |||
| Child-Pugh (A) | 0.60 | 0.09–4.14 | 0.326 | |||
| Maximum tumor diameter | 0.98 | 0-2868.11 | < 0.001 | 0.98 | 0-780.26 | 0.001 |
| Number of lesions | 0.86 | 0.10–7.59 | 0.267 | |||
| Tumor Capsule (Presence) | 0.09 | 0-8.39 | 0.022 | 0.12 | 0-5.98 | 0.046 |
| Portal Vein Invasion (Absence) | 0.39 | 0.01–27.53 | 0.030 | 0.70 | 0.16–3.07 | 0.450 |
| Fistula (Absence) | 1.97 | 0.35–11.11 | 0.378 | |||
Note: Data in parentheses are reference level.
Abbreviations: CI, confidence intervals; AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate transaminase.
Figure 3.
Clinical application of the nomogram for predicting treatment response to DEB-TACE in two representative HCC patients. Case 1: A 41-year-old male patient presented with a maximum tumor diameter of 32 mm, absence of a tumor capsule, and a Radscore of −0.141256. (A) The baseline contrast-enhanced CT (CECT) image showed two tumor lesions in the right hepatic lobe (arrows). (B) Based on the patient’s variables, the nomogram predicted a 96.9% probability of treatment response. (C) A follow-up CECT scan after DEB-TACE confirmed complete necrosis of the lesions (arrows), which was evaluated as a complete response (CR). Case 2: A 52-year-old male patient presented with a maximum tumor diameter of 95 mm, absence of a tumor capsule, and a Radscore of 1.029315. (D) The baseline CECT image showed a large lesion in the right hepatic lobe (arrow). (E) The nomogram predicted a 25.0% probability of treatment response. (F) A follow-up CECT scan after DEB-TACE revealed significant enlargement of the lesion (arrow), which was evaluated as progressive disease (PD).
Table 4.
The Specific Performances of Models for Predicting Treatment Response
| Model | Cohort | Criterion | |||
|---|---|---|---|---|---|
| AUC | Accuracy | Sensitivity | Specificity | ||
| Clinical | Training | 0.800 (0.713–0.888) | 0.788 | 0.731 | 0.798 |
| Internal verification | 0.750 (0.605–0.895) | 0.810 | 0.600 | 0.841 | |
| External verification | 0.618 (0.462–0.774) | 0.776 | 0.400 | 0.936 | |
| Radiomics | Training | 0.842 (0.764–0.921) | 0.857 | 0.654 | 0.890 |
| Internal verification | 0.809 (0.626–0.991) | 0.861 | 0.700 | 0.884 | |
| External verification | 0.742 (0.619–0.864) | 0.701 | 0.900 | 0.617 | |
| Nomogram | Training | 0.915 (0.867–0.962) | 0.820 | 0.885 | 0.810 |
| Internal verification | 0.816 (0.659–0.973) | 0.772 | 0.800 | 0.768 | |
| External verification | 0.785 (0.676–0.894) | 0.731 | 0.850 | 0.681 | |
| BCLC | Training | 0.571 (0.469–0.673) | 0.354 | 0.846 | 0.276 |
| Internal verification | 0.713 (0.601–0.825) | 0.595 | 0.800 | 0.565 | |
| External verification | 0.612 (0.460–0.765) | 0.701 | 0.500 | 0.787 | |
| CNLC | Training | 0.630 (0.524–0.737) | 0.640 | 0.577 | 0.650 |
| Internal verification | 0.794 (0.694–0.894) | 0.722 | 0.800 | 0.710 | |
| External verification | 0.647 (0.497–0.798) | 0.746 | 0.450 | 0.872 | |
| mHAP-II | Training | 0.603 (0.495–0.711) | 0.275 | 1.000 | 0.156 |
| Internal verification | 0.654 (0.483–0.824) | 0.759 | 0.400 | 0.812 | |
| External verification | 0.639 (0.486–0.792) | 0.716 | 0.450 | 0.830 | |
Abbreviations: AUC, area under the receiver operating characteristic curve; BCLC, Barcelona Clinic Liver Cancer staging system; CNLC, China Liver Cancer Staging; mHAP-II, modified Hepatoma Arterial-embolization Prognostic II score.
Figure 4.
Receiver operating characteristic curves for all models in the training (A), validation cohorts (B), and external validation cohort (C).
Table 5.
Results of DeLong Test for ROC Comparison
| Model | Cohort | Model | |||||
|---|---|---|---|---|---|---|---|
| Clinical | Radiomics | Nomogram | BCLC | CNLC | mHAP-II | ||
| Clinical | Training | / | P=0.490 | P=0.004 | P=0.001 | P=0.016 | P=0.006 |
| Internal verification | / | P=0.630 | P=0.528 | P=0.693 | P=0.624 | P=0.400 | |
| External verification | / | P=0.238 | P=0.068 | P=0.958 | P=0.792 | P=0.853 | |
| Radiomics | Training | P=0.490 | / | P=0.010 | P<0.001 | P=0.002 | P<0.001 |
| Internal verification | P=0.630 | / | P=0.826 | P=0.383 | P=0.892 | P=0.225 | |
| External verification | P=0.238 | / | P=0.096 | P=0.198 | P=0.343 | P=0.307 | |
| Nomogram | Training | P=0.004 | P=0.010 | / | P<0.001 | P<0.001 | P<0.001 |
| Internal verification | P=0.528 | P=0.826 | / | P=0.298 | P=0.820 | P=0.172 | |
| External verification | P=0.068 | P=0.096 | / | P=0.035 | P=0.149 | P=0.131 | |
| BCLC | Training | P=0.001 | P<0.001 | P<0.001 | / | P=0.047 | P=0.695 |
| Internal verification | P=0.693 | P=0.383 | P=0.298 | / | P=0.001 | P=0.550 | |
| External verification | P=0.958 | P=0.198 | P=0.035 | / | P=0.294 | P=0.810 | |
| CNLC | Training | P=0.016 | P=0.002 | P<0.001 | P=0.047 | / | P=0.725 |
| Internal verification | P=0.624 | P=0.892 | P=0.820 | P=0.001 | / | P=0.142 | |
| External verification | P=0.792 | P=0.343 | P=0.149 | P=0.294 | / | P=0.938 | |
| mHAP-II | Training | P=0.006 | P<0.001 | P<0.001 | P=0.695 | P=0.725 | / |
| Internal verification | P=0.400 | P=0.225 | P=0.172 | P=0.550 | P=0.142 | / | |
| External verification | P=0.853 | P=0.307 | P=0.131 | P=0.810 | P=0.938 | / | |
Note: Data are P values calculated by DeLong test for ROC comparison between different models in the three cohorts.
Abbreviations: ROC, receiver operating characteristic curve; BCLC, Barcelona Clinic Liver Cancer staging system; CNLC, China Liver Cancer Staging; mHAP-II, modified Hepatoma Arterial-embolization Prognostic II score.
Calibration curves demonstrated good consistency between the nomogram’s predictions and actual outcomes across all cohorts (Figure 5A). The Hosmer-Lemeshow goodness-of-fit test showed no statistically significant differences between predicted and actual results (all P>0.05), indicating good model fit. DCA revealed that the nomogram could provide clinical net benefit for HCC patients at risk thresholds between 0.20 and 0.99 in the training and internal validation cohorts, and between 0.85 and 0.99 in the external validation cohort (Figure 5B).
Figure 5.
Calibration curves (A) and decision curve analysis (B) for the nomogram.
Prognostic Value of the Nomogram
The median follow-up for the entire cohorts was 12.1 months (range: 1.1–60.8 months). Kaplan-Meier analysis confirmed that both the actual mRECIST response (Figure 6A–C) and the nomogram-predicted response (Figure 6D–F) provided significant prognostic stratification for OS.
Figure 6.
Kaplan-Meier curves for overall survival (OS) in patients with HCC. (A–C) OS stratified by the actual therapeutic response defined by mRECIST (CR/PR vs SD/PD) in (A) the training cohort, (B) the internal validation cohort, and (C) the external validation cohort. (D–F) OS stratified by the nomogram-predicted response in (D) the training cohort, (E) the internal validation cohort, and (F) the external validation cohort. All survival curve comparisons were performed using the Log rank test.
Abbreviations: OS, overall survival; CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease.
Focusing on the nomogram’s performance, patients in the predicted CR/PR group had a significantly longer median OS than those in the predicted SD/PD group in the training cohort (42.37 vs 10.79 months; P<0.001). This robust prognostic stratification was consistently validated in the internal validation cohort (median OS: 33.55 vs 18.98 months; P=0.003) and the external validation cohort, where the median OS for the predicted CR/PR group was not reached, while it was 11.23 months for the predicted SD/PD group (P=0.011).
Discussion
In recent years, DEB-TACE has increasingly been applied for treatment of unresectable HCC due to enhanced drug delivery to tumors.10 The efficacy of DEB-TACE varied considerably among HCC patients because of the heterogeneity in tumor burden and characteristics.5 In this study, a novel nomogram integrating maximum tumor diameter, tumor capsule status, and Radscore was constructed to predict tumor response to DEB-TACE in HCC patients. This nomogram exhibited excellent predictive performance in both the training and validation cohorts with the highest AUC compared with other models, and it also showed good calibration and significant clinical net benefit. In addition, an online interactive nomogram tool was developed based on this model, enabling rapid preoperative assessment and facilitating direct application by clinicians.
BCLC, CNLC, and mHAP-II are three classic staging/scoring systems commonly used for HCC assessment, but they have inherent limitations in predicting DEB-TACE efficacy. Notably, these three models differ significantly in their study populations: BCLC staging and mHAP-II scoring are mainly constructed based on Western populations, where HCC is predominantly caused by alcoholic liver disease and hepatitis C virus infection; in contrast, CNLC staging, independently developed in China, is specifically designed for HCC related to hepatitis B virus infection.11,12 Regardless of their population differences, all three models are not optimized for DEB-TACE and lack radiomics information that can reflect the microscopic heterogeneity of tumors, which limits their predictive performance. The nomogram developed in this study offers several distinct advantages. First, it incorporates radiomics features to quantify tumor microscopic heterogeneity, thereby addressing a critical limitation of conventional staging and scoring systems that rely solely on macroscopic clinical parameters. Quantitative comparisons demonstrated that the nomogram achieved higher AUC values than the three classical models across all cohorts. Second, the study cohort comprised patients with diverse etiologies from multiple medical centers, which confers robust generalizability to the model. This nomogram is applicable not only for predicting tumor response to DEB-TACE, but also for risk stratification in patients undergoing alternative treatment regimens, thereby informing personalized therapeutic strategies. Finally, to ensure pre-treatment applicability, operative parameters such as DEB size and loaded drug were deliberately excluded from the model, enabling evaluation prior to intervention.
Multivariate logistic regression analysis revealed that tumor diameter >5 cm and absence of tumor capsule were independent risk factors for failure to achieve ORR following DEB-TACE. Tumor size is a well-established prognostic indicator for HCC, reflecting tumor burden and incorporated into various staging systems.13,14 In the management of large tumors, the efficacy of DEB-TACE may be constrained. Alan et al found that the objective response rate of DEB-TACE exhibited a significant decline when the tumor diameter exceeded 7 cm, potentially attributable to incomplete embolization or the formation of collateral vessels15 Furthermore, maximal tumor diameter has been identified as a critical determinant of both tumor response and survival in patients with large HCC undergoing DEB-TACE.5 These findings suggested that DEB-TACE remains effective for small tumors, and its efficacy may not be as pronounced as it is for large tumors. A significant correlation between the presence of tumor capsule and the tumor response rate following DEB-TACE treatment has been identified.16 Moreover, the tumor capsule may impact tumor progression and survival. Research has demonstrated that the presence of a tumor capsule was associated with extended progression-free survival (PFS), potentially due to the capsule as a barrier to limit the metastasis of tumor cells so as to inhibit tumor progression.17
In recent years, radiomics analysis has garnered increased attention for its pivotal role in providing critical insights into tumor diagnosis, treatment guidance and efficacy and prognosis assessment, with a multitude of quantitative features from imaging data.18–20 Furthermore, the integration of clinical and radiological features enabled the development of comprehensive predictive models to improve the precision of predictions. Specifically, radiomics analysis has been applied to forecast the outcomes of DEB-TACE by preoperative CT or magnetic resonance imaging (MRI).21,22 One study proposed a model, based on preoperative CT features of 50 patients, effectively predicted the response and survival outcomes of patients with HCC after DEB-TACE treatment without external validation cohort. In the present study, however, there were 268 and 67 patients as the training cohort, internal validation and external validation cohorts, respectively.23 Furthermore, another study developed a predictive model for tumor response in HCC patients after TACE through the analysis of preoperative MRI images, which exhibited strong predictive capabilities in both training and validation cohorts.7 The treatment in above study did not yet distinguish between cTACE and DEB-TACE.
Furthermore, radiomics analysis enables in-depth characterization of the biological mechanisms underlying tumor heterogeneity and treatment response. In this study, a total of 15 radiomic features were selected, comprising one first-order feature, three texture features, and eleven wavelet features. Among these, original_glcm_ClusterTendency and wavelet-HHL_glcm_ClusterShade emerged as the two most heavily weighted positive predictors. Specifically, ClusterTendency characterizes the spatial aggregation of pixel intensities; elevated values suggest a clustered tumor cell distribution with relatively concentrated vascularity, which facilitates the precise deposition of embolic agents within the tumor core. ClusterShade quantifies the asymmetry in the gray-level distribution, with higher values indicating the presence of distinct low-attenuation regions (such as necrosis or cystic changes). This finding signifies heightened sensitivity to ischemia and hypoxia, and thus confers greater susceptibility to the embolic effects of DEB-TACE. Previous studies have employed radiomics analysis to identify specific imaging features associated with enrichment of immune-related pathways following TACE treatment in HCC, further substantiating the intrinsic relationship between radiomic features and tumor biological behavior.24 Collectively, radiomics analysis demonstrates substantial potential in evaluating tumor response to DEB-TACE. Integration of clinical indicators with radiomic signatures enables comprehensive assessment of tumor status, enhances predictive accuracy, and thereby provides a more robust foundation for formulating individualized therapeutic strategies.
While our proposed model demonstrated improved AUC values in the external validation cohort, the DeLong test yielded p-values of 0.068 and 0.096 of the external validation cohort. Although these values did not reach the conventional significance level of 0.05, their proximity to this threshold suggests a potential trend towards improved predictive performance that may be clinically meaningful. This borderline significance could be attributed to the limited sample size of the external validation cohort, which may reduce the statistical power of the DeLong test. Therefore, while we cannot definitively claim statistical superiority, the consistent trend observed warrants further investigation in larger, multi-center studies to validate these preliminary findings.
There were several limitations in this study. (1) The retrospective design of the study introduced a potential risk of selection bias. (2) The primary etiology of liver cirrhosis among the patients was hepatitis B infection, whereas chronic hepatitis C and alcoholic hepatitis are more prevalent in western countries. Therefore, it is essential to validate our nomogram within western populations in future research. (3) Our nomogram model was developed to assess the tumor response and to identify the patients with HCC suitable to receive DEB-TACE. Considering the increasing application of combination therapies including DEB-TACE and systemic therapy/immunotherapy in clinical practice, it is important to develop another model to predict the efficacy of combination therapy. (4) This was a dual-center study involving two different CT scanners, which could introduce variability and potentially affect the robustness of the radiomics features. To mitigate this, we implemented image preprocessing to reduce the impact of the varied scanners. (5) To mitigate the limitation of class imbalance while preserving the real-world data distribution, we used the robust AUC metric for evaluation instead of applying data resampling. (6) The performance of our model decreased in the external validation cohort compared to the training cohort, which suggests a degree of overfitting and indicates that its generalizability could be further improved.
Conclusions
In summary, we proposed a reliable nomogram with superior predictive capability for assessing tumor response to DEB-TACE with clinic and radiomics data. This nomogram model has the potential to identify the patients who are suitable to undergo DEB-TACE, and thereby developing treatment and follow-up strategies.
Funding Statement
This study has received funding by Henan Province Natural Science Foundation key science fund project (232300421120), Central Plains Science and Technology Innovation Leadership Talent Program (244200510020) and Henan Provincial Key R&D Program (241111311000).
Abbreviations
HCC, Hepatocellular carcinoma; TACE, transarterial chemoembolization; BCLC, the Barcelona Clinic Liver Cancer staging system; DEB-TACE, drug-eluting bead TACE; ORR, objective response rate; DCR, disease control rate.
Data Sharing Statement
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Disclosure
The authors report no conflicts of interest in this work.
References
- 1.Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17–14. doi: 10.3322/caac.21763 [DOI] [PubMed] [Google Scholar]
- 2.Diao X, Guo C, Jin Y, et al. Cancer situation in China: an analysis based on the global epidemiological data released in 2024. Cancer Commun. 2025;45(2):178–197. doi: 10.1002/cac2.12627 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Reig M, Forner A, Rimola J, et al. BCLC strategy for prognosis prediction and treatment recommendation: the 2022 update. J Hepatol. 2022;76(3):681–693. doi: 10.1016/j.jhep.2021.11.018 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Shi Z, Wang D, Kang T, et al. Comparison of CalliSpheres® microspheres drug-eluting beads and conventional transarterial chemoembolization in hepatocellular carcinoma patients: a randomized controlled trial. Radiol Oncol. 2023;57(1):70–79. doi: 10.2478/raon-2023-0001 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Ji K, Niu J, Zhang C, et al. Systemic inflammation-based staging system for hepatocellular carcinoma after drug-eluting beads transarterial chemoembolization: a multicenter study. Acad Radiol. 2025;32(2):776–786. doi: 10.1016/j.acra.2024.08.032 [DOI] [PubMed] [Google Scholar]
- 6.Kadalayil L, Benini R, Pallan L, et al. A simple prognostic scoring system for patients receiving transarterial embolisation for hepatocellular cancer. Ann Oncol. 2013;24(10):2565–2570. doi: 10.1093/annonc/mdt247 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Kong C, Zhao Z, Chen W, et al. Prediction of tumor response via a pretreatment MRI radiomics-based nomogram in HCC treated with TACE. Eur Radiol. 2021;31(10):7500–7511. doi: 10.1007/s00330-021-07910-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Wang Q, Nilsson H, Xu K, et al. Exploring tumor heterogeneity in colorectal liver metastases by imaging: unsupervised machine learning of preoperative CT radiomics features for prognostic stratification. Eur J Radiol. 2024;175:111459. doi: 10.1016/j.ejrad.2024.111459 [DOI] [PubMed] [Google Scholar]
- 9.He X, Li K, Wei R, et al. A multitask deep learning radiomics model for predicting the macrotrabecular-massive subtype and prognosis of hepatocellular carcinoma after hepatic arterial infusion chemotherapy. Radiol Med. 2023;128(12):1508–1520. doi: 10.1007/s11547-023-01719-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Ji K, Shi Y, Liang Z, et al. Lipiodol combined with drug-eluting beads versus drug-eluting beads alone for transarterial chemoembolization of hepatocellular carcinoma: a multicenter study. Acad Radiol. 2024;31(12):4912–4922. doi: 10.1016/j.acra.2024.05.033 [DOI] [PubMed] [Google Scholar]
- 11.Peisen F, Maurer M, Grosse U, et al. Predictive performance of the mHAP-II score in a real-life western cohort with hepatocellular carcinoma following trans-arterial chemoembolisation with drug-eluting beads (DEB-TACE). Eur Radiol. 2020;30(7):3782–3792. doi: 10.1007/s00330-020-06734-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Mahringer-Kunz A, Weinmann A, Schmidtmann I, et al. Validation of the SNACOR clinical scoring system after transarterial chemoembolisation in patients with hepatocellular carcinoma. BMC Cancer. 2018;18(1):489. doi: 10.1186/s12885-018-4407-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Wang Q, Xia D, Bai W, et al. Development of a prognostic score for recommended TACE candidates with hepatocellular carcinoma: a multicentre observational study. J Hepatol. 2019;70(5):893–903. doi: 10.1016/j.jhep.2019.01.013 [DOI] [PubMed] [Google Scholar]
- 14.Ji K, Zhu H, Zhang C, et al. Nomogram-based prognostic stratification for patients with large hepatocellular carcinoma: a population study of SEER database and a Chinese cohort. J Gastrointest Oncol. 2024;15(5):2201–2215. doi: 10.21037/jgo-24-288 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Alan AM, Alan O, Asadov R, et al. Evaluation of the effectiveness of drug-eluting transarterial chemoebolization in hepatocellular carcinoma. Hepatol Forum. 2023;4(2):53–60. doi: 10.14744/hf.2022.2022.0048 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Hatakeyama K, Tozawa T, Wada Y, et al. Tumor diameter and enhancing capsule, as well as previous interventional treatments, as potential predictors of vascular lake phenomenon in hepatocellular carcinoma patients treated with drug-eluting beads transarterial chemoembolization. J Clin Imaging Sci. 2024;14:29. doi: 10.25259/JCIS_66_2024 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Kim MS, Oh JS, Chun HJ, et al. Comparison of small-sized (70-150 mum) and intermediate-sized (100-300 mum) drug-eluting embolics for transarterial chemoembolization of small hepatocellular carcinomas (</=3 cm). J Vasc Interv Radiol. 2024;35(11):1626–1634e3. doi: 10.1016/j.jvir.2024.06.032 [DOI] [PubMed] [Google Scholar]
- 18.Zhan PC, Yang T, Zhang Y, et al. Radiomics using CT images for preoperative prediction of lymph node metastasis in perihilar cholangiocarcinoma: a multi-centric study. Eur Radiol. 2024;34(2):1280–1291. doi: 10.1007/s00330-023-10108-1 [DOI] [PubMed] [Google Scholar]
- 19.Li HJ, Qiu ZB, Wang MM, et al. Radiomics-based support vector machine distinguishes molecular events driving the progression of lung adenocarcinoma. J Thorac Oncol. 2025;20(1):52–64. doi: 10.1016/j.jtho.2024.09.1431 [DOI] [PubMed] [Google Scholar]
- 20.Yu F, Yang M, He C, et al. CT radiomics combined with clinical and radiological factors predict hematoma expansion in hypertensive intracerebral hemorrhage. Eur Radiol. 2025;35(1):6–19. doi: 10.1007/s00330-024-10921-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Zhang X, He Z, Zhang Y, et al. Prediction of initial objective response to drug-eluting beads transcatheter arterial chemoembolization for hepatocellular carcinoma using CT radiomics-based machine learning model. Front Pharmacol. 2024;15:1315732. doi: 10.3389/fphar.2024.1315732 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Kuang D, Zhang N, Zhang M, et al. Correlation between magnetic resonance images of peritumor margin enhancement and prognosis in hepatocellular carcinoma after drug-eluting bead transcatheter arterial chemoembolization. Front Oncol. 2023;13:957710. doi: 10.3389/fonc.2023.957710 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Tipaldi MA, Ronconi E, Lucertini E, et al. Hepatocellular carcinoma drug-eluting bead transarterial chemoembolization (DEB-TACE): outcome analysis using a model based on pre-treatment CT texture features. Diagnostics. 2021;11(6). doi: 10.3390/diagnostics11060956 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Zhang X, Teng X, Zhang J, et al. Enhancing pathological complete response prediction in breast cancer: the role of dynamic characterization of DCE-MRI and its association with tumor heterogeneity. Breast Cancer Res. 2024;26(1):77. doi: 10.1186/s13058-024-01836-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.