Shangman Yang1,2,*, Zerong Huang1,*, Wei Zheng1, Lingling Li1, Yao Xue1,3, Ruhai Zou1, Chuan Peng1

1Department of Ultrasound, Sun Yat-Sen University Cancer Center, Guangzhou, Guangdong, People’s Republic of China; 2Department of Ultrasound Medicine, The Affiliated Hospital of Guizhou Medical University, Guiyang, Guizhou, People’s Republic of China; 3Department of Ultrasound, Zhongshan Hospital of Traditional Chinese Medicine, Zhongshan, Guangdong, People’s Republic of China

Correspondence: Chuan Peng, Department of Ultrasound, Sun Yat-Sen University Cancer Center, 651 Dongfeng Road East, Guangzhou, 510060, People’s Republic of China, Tel +86 020 87343212, Fax +86 020-87343212, Email [email protected]; Ruhai Zou, Department of Ultrasound, Sun Yat-Sen University Cancer Center, 651 Dongfeng Road East, Guangzhou, 510060, People’s Republic of China, Tel +86 020-87343212, Fax +86 020-87343212, Email [email protected]

Objective: To develop and validate a noninvasive radiomics approach based on Sonazoid contrast-enhanced ultrasound (CEUS) for preoperative prediction of Glypican-3 (GPC3) status in solitary hepatocellular carcinoma (HCC) patients and to evaluate its association with early recurrence-free survival (RFS).
Materials and Methods: This diagnostic accuracy study retrospectively included 278 solitary HCC patients at a grade A Tertiary Hospital (2020.3– 2024.1). Radiomics features were extracted from CEUS images, and robust feature selection was performed using minimum redundancy-maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO). A logistic-regression radiomics model was trained and internally validated using a pathologic reference standard. A hybrid model combining the radiomics signature with significant clinical variables was then constructed. Model performance was assessed by the area under the receiver operating characteristic curve (AUC). RFS was compared using the log‑rank test.
Results: Among 278 patients (median age, 57 years; IQR, 49– 64 years; 232 males), alpha-fetoprotein (AFP) > 200 ng/mL and the radiomics score were independent predictors of GPC3 positivity. The hybrid model integrating CEUS-based radiomics features and AFP achieved AUCs of 0.81 (95% CI: 0.74– 0.89) and 0.78 (95% CI: 0.66– 0.89) in the training and validation sets, respectively, significantly outperforming the radiomics model (AUC 0.74 (95% CI: 0.65– 0.83) and 0.67 (95% CI: 0.52– 0.83), both p < 0.01). The radiomics score was not significantly associated with early RFS (p > 0.05).
Conclusion: A hybrid model integrating CEUS-based radiomics features and AFP, developed in solitary HCC, shows potential for non-invasive preoperative prediction of GPC3 positive expression in HCC, which may facilitate personalized preoperative treatment planning for these patients. However, it did not predict early recurrence-free survival, and its prognostic value was not demonstrated in this study.

Keywords: hepatocellular carcinoma, glypican-3, contrast-enhanced ultrasound, radiomics

Introduction

According to the 2022 global cancer statistics, primary liver cancer ranks sixth in incidence and third in mortality among malignant tumors worldwide, and hepatocellular carcinoma (HCC) accounts for approximately 80% of all primary liver cancer cases.1 Despite advances in diagnostic techniques,2 the majority of HCC patients are diagnosed at an intermediate or advanced stage, where treatment options are non-curative and limited.3,4 The development of more effective systemic therapies, along with predictive biomarkers that enable personalized and cost-effective treatment stratification, has become a major focus of current research. As a target for immunotherapeutic strategies, Glypican-3 (GPC3) demonstrated promising results over the past decade.5

GPC3 has emerged as a pivotal biomarker and therapeutic target in HCC. It is a glycosylphosphatidylinositol-anchored cell-surface proteoglycan that regulates proliferation, migration, differentiation, and survival through Wnt signal, insulin-like growth factor-2, and related pathways.6 It is markedly upregulated in several malignancies, including HCC, ovarian cancer, and melanoma.7 In HCC, it has diagnostic, therapeutic, and prognostic relevance. Tissue GPC3 expression accurately distinguishes HCC from other liver diseases,8–10 whereas serum levels vary considerably across studies and populations, limiting immediate clinical utility and requiring further validation.11–14 Therapeutically, GPC3-targeted therapeutic methods have shown promising results, including GPC3 vaccine, anti-GPC3 immunotoxin, combined therapy with immune checkpoint blockades, and chimeric antigen receptor T or NK cells.15 Prognostically, GPC3 expression correlates with poor outcomes in some studies16–19 but not others.20–22 Thus, preoperative treatment planning could benefit from reliable prediction of GPC3 status.

Preoperative prediction of GPC3 is challenging because the current reference standard remains immunohistochemistry of postoperative specimens. Previous studies have shown that CT/MRI features can predict GPC3 status (AUC, 0.79–0.96).21,23–29 Compared with CT/MRI, contrast-enhanced ultrasound (CEUS) offers advantages such as high spatiotemporal resolution, absence of ionizing radiation, low cost, and real-time visualization. Among CEUS contrast agents, Sonovue has been used to identify predictors for GPC3 positive HCC.30–33 Sonovue cannot cross the vascular endothelium and is assumed to generate pure vascular images,34 its short duration of contrast enhancement hinders prolonged observation.35 The contrast agent perfluorobutane (Sonazoid; GE Healthcare) can be specifically phagocytosed by macrophages in the liver, thus allowing additional lesion characterization in the unique Kupffer phase.36,37 Given that GPC3 has been shown to be involved in the recruitment of M2-polarized tumor-associated macrophages in HCC,38 we hypothesize that Sonazoid contrast-enhanced ultrasound (CEUS) may serve as a predictor for GPC3 status. Radiomics is a newly emerging form of imaging analysis that bridges medical imaging and personalized medicine through the high-throughput extraction of quantitative features from standard-of-care medical imaging.39 Previous CEUS-derived radiomics studies have provided insights into multiple aspects of HCC, including microvascular invasion, Ki-67 expression and CK19 expression.40–44 However, the potential of Sonazoid CEUS radiomics to non-invasively predict GPC3 expression remains unexplored.

Accordingly, this study aimed to develop and validate a non-invasive radiomics approach for predicting GPC3 status based on preoperative Sonazoid CEUS and to evaluate its association with early recurrence-free survival (RFS).

Materials and Methods Patients

This single center, retrospective study was approved by the Institutional Review Board (B2025-070-01), and the requirement for informed consent was waived because it involved only review of existing medical records and anonymized imaging data. All patient data were anonymized and handled with strict confidentiality. Consecutive patients with HCC who underwent preoperative Sonazoid CEUS within one month prior to hepatectomy between March 2020 and January 2024 were identified through the Yiduyun big data intelligence platform at our center. Inclusion criteria were as follows: (a) pathologically confirmed HCC, (b) availability of GPC3 immunohistochemistry, (c) preoperative Sonazoid CEUS within one month before hepatectomy, (d) treatment-naïve patients undergoing surgical resection only, (e) a solitary lesion with maximum diameter > 2 cm. Exclusion criteria included: (a) any preoperative antitumor therapy, (b) history of other malignancies or presence of extrahepatic metastasis, (c) incomplete clinical data, (d) poor image quality (eg, motion/registration artifacts), (e) macroscopic vascular invasion or lymph node metastasis. A final total of 278 patients were enrolled. Patients were then randomly divided into a training cohort (n = 195, 70%) and a validation cohort (n = 83, 30%). The patient selection flow diagram is shown in Figure 1.

Figure 1 Patient selection flowchart. From 466 patients with resected HCC who had preoperative Sonazoid contrast-enhanced ultrasound (CEUS), available GPC3 immunohistochemistry (IHC), and a solitary lesion (March 2020–January 2024), 188 were excluded for the reasons shown, leaving 278 patients for analysis (GPC3 positive, n = 227; GPC3 negative, n = 51).

Abbreviations: CEUS, contrast-enhanced ultrasound; IHC, immunohistochemistry; GPC3, glypican-3.

CEUS Acquisition Protocols

Sonazoid CEUS examinations were performed with an Acuson Sequoia 512 system (Siemens Medical Solutions). A bolus of Sonazoid (GE Healthcare) at 0.015 mL/kg was injected via the antecubital vein, immediately followed by a 5 mL saline flush; dynamic imaging was then initiated. Real-time imaging was recorded, with continuously monitoring of lesions and surrounding hepatic parenchyma for at least 80 seconds. After 80 seconds, the lesions were intermittently scanned and recorded for 5 minutes in the vascular phase (ie, arterial phase, portal venous phase, and delayed phase), and the lesions were re-evaluated in the Kupffer phase from 10 to 30 minutes.

Clinical Data and Image Analysis

Clinical parameters, including sex, age, hepatitis virus infection, liver cirrhosis, preoperative hematological parameters (eg, alpha-fetoprotein [AFP], alanine aminotransferase [ALT]), were collected from electronic medical records.

One radiologist (C.P, with over 10 years of experience) who was blinded to the results of pathologic evaluation and other imaging findings independently reviewed all US and CEUS images. Any discrepancies in interpretation were resolved by consensus with a second senior radiologist (R.H.Z, with over 20 years of experience). The interpretation of CEUS images was conducted to the CEUS LI-RADS45 and in reference to the criteria described in a Radiology publication34 (detailed in Appendix S1).

Pathologic GPC3 expression status was obtained through pathologic evaluation from electronic health records. Following established criteria,46 GPC3 positive expression (GPC3+) was defined as that immunoreactive cells were greater than (10%) of tumor cells, otherwise it was defined as negative (GPC3-).

ROI Delineation, Radiomics Feature Extraction, and Selection

For each target lesion, five images were selected based on the largest cross-sectional area and clear boundaries in US, and CEUS (arterial, portal venous, delayed and Kupffer). A radiologist (S.M.Y with 3 years of experience), blinded to pathology, manually delineated the regions of interest (ROIs) using the open-source annotation tool LabelMe47 (Appendix S2).

Image preprocessing and radiomics features extraction were performed in Python. A total of 102 radiomics features (including 9 shape features, 18 first-order features and 75 second-order features) were extracted per image (Appendix S3), yielding 510 features per tumor.

The extracted radiomics features were subjected to Z-score normalization. In the training set, the minimum redundancy maximum relevance algorithm (mRMR) was used to select the top 20 features, followed by the least absolute shrinkage and selection operator regression (LASSO) with ten-fold cross-validation method to identify the key radiomics features. A radiomics score (radscore) was calculated based on the selected features.

Reproducibility Analysis

To assess inter- and intra-observer reproducibility, 30 cases were randomly selected. Intra-observer analysis was performed by the same radiologist after a washout period. Inter-observer analysis was performed independently by a second radiologist. The LASSO-selected radiomics features all demonstrated good reproducibility, with intra-observer ICCs > 0.75 and inter-observer ICCs > 0.75.

Model Building and Testing

A radiomics model was developed using the radscore to calculate the predictive probability for each sample by logistic regression. We analyzed clinical parameters, imaging features and radscore via uni- and multivariable analysis. The variables associated with GPC3 in univariable logistic analysis (p < 0.1 were incorporated into a multivariable analysis logistic regression. A hybrid model was developed integrating radscore and significant clinical features (p < 0.05 in the multivariable analysis).

We calculated the area under the receiver operating curve (AUC) for different models to evaluate their predictive performance for GPC3 and compared them using Delong’s test. Model calibration was assessed using calibration curves and clinical utility was evaluated via decision curve analysis. The cutoff values for the models were determined based on the maximum Youden’s index. We also compared the accuracy (ACC), sensitivity (SEN), and specificity (SPE) across the different models.

Outcome Prediction

Patients were divided into high-score and low-score groups based on the Youden index of the radiomics score and hybrid score in the cohort which patients followed for more than 2 years. The log‑rank test was used to compare early recurrence-free survival (RFS) between low or high score group. The follow-up strategy is detailed in Appendix S4.

Statistical Analysis

This study was reported in accordance with the Transparent Reporting of a multivariable prediction model for individual Prognosis or Diagnosis (TRIPOD).48 Statistical analysis was performed with R software (version 4.4.2) and Python (version 3.8.2). Due to non-normal distribution, continuous variables were reported as medians with interquartile ranges (IQRs) and compared using the Mann–Whitney U-test; categorical variables were reported as frequencies and percentages, and compared using chi-square test or Fisher’s exact test. Image annotation of the ROIs was performed using the open-source tool LabelMe. Radiomics feature extraction was carried out in Python, while the selection of key radiomics features was implemented in R.

Univariable and multivariable logistic regressions were used to identify risk factors with odds ratio (OR) and 95% confidence intervals (CIs). The model performance evaluation included the calculation and comparison of sensitivity, specificity, accuracy, and AUC. Two-sided p < 0.05 indicated statistical significance. Calibration plots were used to compare the predicted values with the observed values and decision curve analysis was used to evaluate the net benefits at different threshold probabilities. To address class imbalance (81.7% GPC3-positive vs. 18.3% GPC3-negative), class weighting was applied in the logistic regression model, assigning higher weights to GPC3-negative cases. The weighted results were consistent with the unweighted analysis, confirming robustness.

Results Characteristics of the Study Sample

Our study initially screened 466 eligible participants with HCC, 62 were excluded from this study because of preoperative treatment, 53 were excluded because of other malignancies or extrahepatic metastasis, 8 were excluded because of incomplete clinical data, 56 were excluded because of poor image quality and 9 were excluded because of macroscopic vascular invasion or lymph node metastasis (detailed in Figure 1). A final total of 278 patients (median age, 57 years; IQR, 49–64 years; 232 males) with solitary HCC were enrolled in the study, 227 (81.7%) with GPC3 positive HCC and 51 (18.3%) with GPC3 negative HCC. Of the 278 patients in the final cohort, 96 (34.5%) were older than 60 years, 232 (83.5%) were males, and 256 of 278 (92.1%) had underlying hepatitis virus infection. The ultrasonographic features were as follows: of the 278 included patients, 120 (43.2%) had a maximum tumor diameter >5 cm, 197 (70.9%) showed hypoechogenicity, 240 (86.3%) exhibited heterogeneous internal structure, 127 (45.7%) had an irregular shape, and 209 (75.2%) had an ill-defined margin. All lesions (100%) exhibited arterial phase hyperenhancement (APHE), with early washout (<60 s) observed in 157 (56.5%). In the portal venous phase, 212 (76.3%) showed hypoenhancement, which was also observed in 271 (97.5%) during the delayed phase. In the Kupffer phase, 204 (73.4%) demonstrated mild washout. The baseline characteristics between training and validation cohorts are summarized in Table 1 (all p-values > 0.05).

Table 1 Baseline Clinical Characteristics

In the training set, GPC3-positive status was significantly associated with AFP levels higher than 200 ng/mL (73/159 [46%] vs 4/36 [11%]; p < 0.01), a maximum tumor diameter greater than 5 cm (61/159 [38%] vs 11/36 [31%]; p = 0.03), an irregular shape (84/159 [53%] vs 11/36 [31%]; p = 0.02), and hepatitis positive (151/159 [95%] vs 28/36 [78%]; p < 0.01). No other baseline characteristics showed significant differences between GPC3-positive and GPC3-negative groups, as detailed in Table 2.

Table 2 Univariable and Multivariable Logistic Regression Analysis Based on the Training Cohort

Performance of Radiomics Model for GPC3 Prediction

In this study, we initially extracted 510 radiomics features from tumor ROIs across five imaging phases. Through sequential feature selection employing mRMR and LASSO algorithm, seven robust features were identified as key predictive features, including original_gldm_SmallDependenceLowGrayLevelEmphasis, original_glszm_LargeAreaHighGrayLevelEmphasis, original_glszm_GrayLevelNonUniformity, original_shape2D_Elongation, V.original_glrlm_LongRunHighGrayLevelEmphasis, V.original_ngtdm_Contrast, Kupffer.original_glcm_MCC. Based on these selected features, we constructed a radiomics score (radscore) for each patient using the formula in Appendix S5.

Patients who were GPC3-positive exhibited significantly higher radscore compared to those who were GPC3-negative in both cohorts, with both p-values being<0.05 (Figure 2a). The performance of the radiomics model is displayed in Table 3. The AUC of the radiomics model was AUC 0.74 (95% CI:0.65–0.83), with a sensitivity of 71% (113/159), a specificity of 72% (26/36) and an accuracy of 71% (139/195) in the training set. In the validation set, the AUC was 0.67 (95% CI:0.52–0.83), with a sensitivity of 66% (45/68), a specificity of 60% (9/15) and an accuracy of 65% (54/83).

Table 3 Diagnostic Performance of Hybrid and Radiomics Models

Figure 2 Cloud and rain plots comparing Radscore (a) Hybridscore (b) between the GPC3-negative and GPC3-positive groups across the training and validation cohorts. Assessment of models for the ability to predict GPC3 status. Areas under the receiver operating characteristic curve (AUCs) for the (c) training set and the (d) validation set. Red line represents the hybrid model, which includes AFP and radiomics score. Yellow line represents the radiomics model, which includes 7 radiomics features. AUCs are reported with 95% CI in parentheses.

Abbreviation: GPC3, Glypican-3.

Performance of Hybrid Model for GPC3 Prediction

Univariable analysis of the training set identified seven variables significantly associated with GPC3 positive status: AFP level > 200 ng/mL, maximum tumor diameter > 5 cm, irregular shape, hepatitis virus infection, cirrhosis, heterogeneous internal structure and the radscore (p < 0.1 for all) (Table 2). Subsequent multivariable logistic regression confirmed that AFP level > 200 ng/mL (OR = 7.973; 95% CI: 2.447–25.976; p < 0.01) and the radscore (OR = 11.051; 95% CI: 2.798–43.644; p < 0.01) were independently associated with GPC3 (Table 2). We constructed Hybridscore based on radscore and AFP for each patient, patients who were GPC3-positive exhibited significantly higher Hybridscore compared to those who were GPC3-negative in both cohorts, with both p-values being < 0.05 (Figure 2b).

Compared to the radiomics model, the hybrid model constructed using aforementioned variables for the prediction of GPC3 status demonstrated higher AUC (0.81 vs 0.74, P=0.032), sensitivity (82% vs 71%, p < 0.01), and accuracy (79% vs. 71%, p = 0.01), while maintaining comparable specificity (69% vs 72%, p > 0.05) in the training set. This performance was consistent in the validation set (AUC: 0.78 vs 0.67, p < 0.01; sensitivity (78% vs 66%, p < 0.01); accuracy (76% vs. 65%, p < 0.01); specificity (67% vs 60%, p > 0.05)). Complete performance metrics are detailed in Table 3. The receiver operating characteristic curves for the training and validation sets are presented in Figure 2c and d. The variables of the hybrid model were integrated into a nomogram to provide individualized probability estimates and visualize the weight of individual characteristics (Figure 3). Figures 4 and 5 show representative cases from the GPC3-positive and GPC3-negative groups.

Figure 3 Nomogram of the hybrid model for predicting GPC3 in patients with HCC.

Figure 4 Imaging findings in a male patient over 60 years without hepatitis. The patient’s alpha-fetoprotein (AFP) level was < 200 ng/L. Radiomic analysis yielded a score (Radscore) of 0.628, and a hybrid model predicted a low probability (<30%) of glypican-3 (GPC3) positivity, which was consistent with the final pathological diagnosis of GPC3-negative. The encircled areas indicate the regions of interest (ROI). (a) grayscale ultrasound. (b) arterial phase. (c) portal venous phase. (d) delayed phase. (e) kupffer phase.

Figure 5 Imaging findings in a male patient under 60 years with hepatitis. The patient presented with an alpha-fetoprotein (AFP) level >200 ng/L. Radiomic analysis yielded a score (Radscore) of 2.14, and a hybrid model predicted a high probability (>90%) glypican-3 (GPC3) positivity which was consistent with the final pathological diagnosis of GPC3-positive. The encircled areas indicate the regions of interest (ROI). (a) grayscale ultrasound. (b) arterial phase. (c) portal venous phase. (d) delayed phase. (e) kupffer phase.

The calibration curves also indicated good performance between model-predicted GPC3 probability and the actual GPC3 probability for both groups (all Brier scores < 0.25) (Figure 6a and b). Decision curve analysis showed that the hybrid model provided a larger net benefit across the range of reasonable threshold probabilities than the radiomics model (Figure 6c and d).

Figure 6 The calibration curve in the (a) training and the (b) validation sets. The decision curve in the (c) training and the (d) validation sets.

Survival Analysis

Among the 278 patients, 103 were excluded from the early recurrence-free survival (RFS) analysis because they were either lost to follow-up or had follow-up durations shorter than 2 years at the time of data cutoff (January 2026). Thus, a subgroup of 175 patients (62.9%) with over two years of follow-up (median: 27.6 months) was available for survival analysis, comprising 138 (78.9%) GPC3-positive and 37 (21.1%) GPC3-negative HCC cases. Tumor recurrence occurred in 76 patients (43.4%), of which 69 (39.4%) were early recurrences. Patients were then stratified into high-score and low-score groups based on the Youden index of radiomics score (1.455) and hybrid score (0.726). The 6-, 12-, 18-, 24-month cumulative PFS rate was 91.3%, 73.2%, 65.9%, 59.4% in the GPC3-positive group vs 100.0%, 86.5%, 75.7%, 64.9% in the GPC3-negative group, respectively. No significant difference in RFS was observed between patients stratified by pathological GPC3 status (P = 0.48). Similarly, risk stratification based on the radiomics model (P = 0.11) or the hybrid model (P = 0.93) also failed to show a significant association with RFS (Figure 7).

Figure 7 Kaplan-Meier curves of early recurrence-free survival of radiomics model (a), hybrid model (b), and GPC3 status (c). A high model score is represented by a red line and a low model score is represented by a yellow line. The high and low groups were divided based on the Youden index of the model-driven scores. P values were calculated using a two-sided log‑rank test. The Rad score and hybrid score are derived from the radiomics model and hybrid model, respectively.

Discussion

In this study, we developed and validated the first Sonazoid CEUS based radiomics model for preoperatively predicting GPC3 expression in patients with solitary HCC; subsequently we evaluated its association with early recurrence-free survival (RFS). We found that in multivariable logistic regression analysis, the radscore and AFP were strongly significant predictors of GPC3 status. The hybrid model, which incorporated both AFP and the radscore, achieved significantly higher AUCs than the radiomics model (0.81 vs. 0.74 in the training set and 0.78 vs. 0.67 in the validation set; both p < 0.05). Moreover, the hybrid model showed an improved sensitivity (82% vs 71%, p < 0.01), and an improved accuracy (79% vs 71%, p = 0.01), while maintaining comparable specificity (69% vs 72%, p > 0.05) compared with the radiomics model in the training set. The hybrid model may have potential for integration into clinical practice and therapeutic decision-making. However, neither the model-derived scores nor the GPC3 status significantly stratified HCC patients in terms of early recurrence-free survival (p > 0.05 for all).

Previous studies indicated that CT/MRI features, along with their radiomics models, can predict GPC3 status (AUC, 0.79–0.96).21,23–29 While these results are promising, CT/MRI are associated with higher cost, longer examination time, and exposure to ionization radiation (in case of CT), which may limit their utility for repeated or serial assessments. In contrast, CEUS offers a more accessible, radiation-free, and convenient alternative. Two previous Sonovue CEUS studies attempted to predict GPC3 expression based on visual interpretation of imaging features: one identified AFP level, arterial phase hyperenhancement, asynchronous perfusion as predictive factors (specificity: 60%),30 and another indicated that AFP, younger age and LI-RADS M were independent risk factors for GPC3 positive HCC (AUC:0.759).31 To improve diagnostic performance, we developed and validated a radiomics model based on Sonazoid CEUS for the accurate assessment of GPC3 expression in HCC. Sonazoid provides a unique Kupffer phase linked to GPC3-related macrophage biology, and our quantitative approach captures tumor heterogeneity beyond visual assessment. Li et al recently reported that Sonazoid CEUS-based Kupffer phase radiomics achieved an AUC of 0.70 for predicting GPC3 expression.49 Our hybrid model integrating AFP achieved an AUC of 0.81 in the training set and 0.78 in the validation set, demonstrating its potential as a clinically useful tool.

In the training cohort, univariate analysis revealed significant differences in AFP level > 200 ng/mL, maximum tumor diameter > 5 cm, irregular shape, hepatitis virus infection, cirrhosis, heterogeneous internal structure and the radscore (p < 0.1 for all) between GPC3-positive and -negative groups. Subsequent multivariate analysis confirmed that only radscore and AFP > 200 ng/mL, remained as independent predictors for GPC3 positivity, which is consistent with several previous studies.18,30 AFP serves as a specific diagnostic marker for HCC, and there is a certain correlation between GPC3 expression and AFP in HCC, as AFP and GPC3 may share the transcription factors zinc fingers.50 Though CEUS features and US features showed no statistical significance differing from prior work,30 the radscore held significant predictive value. This underscores a key advantage of radiomics: its ability to extract and quantify subvisual tumor characteristics that elude conventional visual assessment, thereby bridging the gap between macroscopic imaging and microscopic pathology.

In this study, 175 patients underwent a follow-up of over two years, 76 (43.4%) patients experienced tumor recurrence. However, the high versus low scores derived from our radiomics and hybrid models could not stratify patients with early recurrence-free survival (p > 0.05). While previous studies have linked GPC3 positive status to poor prognosis and recurrence,16–19 an MRI-based radiomics showed there was no significant difference in RFS and OS in GPC3-positive and -negative HCC.21 Another study reported that positive expression of tissue GPC3 was associated with a shorter disease-free survival but not overall survival.22 These discrepancies highlight the controversial prognostic role of GPC3. The negative findings in our study may also be limited by the follow-up duration and sample size. Thus, a longer follow-up duration and a larger sample size are required in order to obtain more reliable results for its prognostic evaluation in HCC.

Our study had several limitations. First, this was a single-center retrospective study with class imbalance (81.7% GPC3-positive) and a modest sample of GPC3-negative cases (n=51), which may introduce selection bias and limit generalizability. Patients with preoperative treatment, multiple lesions, or extrahepatic metastasis were excluded to ensure homogeneity, further limiting applicability to broader HCC populations. Second, the analysis was based on static images from key phases rather than the entire dynamic CEUS sequences due to technical challenges, which may not fully capture tumor characteristics and introduces a degree of subjectivity. Third, although post-hoc verification confirms the robustness of the final model, we recognize that pre-selection ICC filtering is methodologically preferable. Finally, the subset of patients with complete two-year follow-up was relatively small (175/278, 62.9%) due to loss to follow-up and incomplete follow-up duration at the time of data cutoff. Extending more centers and the follow-up period is necessary to obtain a larger sample size and more reliable outcomes.

In conclusion, a hybrid model integrating CEUS-based radiomics features and AFP, developed in solitary HCC, shows potential for non-invasive preoperative prediction of GPC3 positive expression in HCC, which may facilitate personalized preoperative treatment planning for these patients. However, it did not predict early recurrence-free survival, and its prognostic value was not demonstrated in this study.

Abbreviations

GPC3, Glypican-3; CEUS, contrast enhanced ultrasound; RFS recurrence-free survival; HCC, hepatocellular carcinoma; mRMR, minimum redundancy–maximum relevance; LASSO, least absolute shrinkage and selection operator; AUC, area under the receiver operating characteristic curve; ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALB, albumin; CRP, C-reactive protein; AFP, α-fetoprotein; PIVKA-II, Protein Induced by Vitamin K Absence or Antagonist – II; APHE, arterial phase hyperenhancement; IQR, interquartile range; OR, odds ratio; CI, confidence interval.

Data Sharing Statement

Raw data are available from the corresponding author, Ruhai Zou on reasonable request.

Ethics Approval

All procedures were approved by the Ethics Committee of Sun Yat-sen University Cancer Center and followed the tenets of the Declaration of Helsinki (Number B2025-070-01).

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

This study was supported by the National Natural Science Foundation of China (No. 82171957).

Disclosure

The authors have no potential conflicts of interest to disclose.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229–14. doi:10.3322/caac.21834

2. Elhendawy M, Abdul-Baki EA, Abd-Elsalam S, et al. MicroRNA signature in hepatocellular carcinoma patients: identification of potential markers. Mol Biol Rep. 2020;47(7):4945–4953. doi:10.1007/s11033-020-05521-4

3. Li D, Lin S, Hong J, Ho M. Immunotherapy for hepatobiliary cancers: emerging targets and translational advances. Adv Cancer Res. 2022;156:415–449. doi:10.1016/bs.acr.2022.01.013

4. Sheta E, El-Kalla F, El-Gharib M, et al. Comparison of single-session transarterial chemoembolization combined with microwave ablation or radiofrequency ablation in the treatment of hepatocellular carcinoma: a randomized-controlled study. Eur J Gastroenterol Hepatol. 2016;28(10):1198–1203. doi:10.1097/MEG.0000000000000688

5. Steffin D, Ghatwai N, Montalbano A, et al. Interleukin-15-armoured GPC3 CAR T cells for patients with solid cancers. Nature. 2025;637(8047):940–946. doi:10.1038/s41586-024-08261-8

6. Zhou F, Shang W, Yu X, Tian J. Glypican-3: a promising biomarker for hepatocellular carcinoma diagnosis and treatment. Med Res Rev. 2018;38(2):741–767. doi:10.1002/med.21455

7. Ho M, Kim H. Glypican-3: a new target for cancer immunotherapy. Eur J Cancer. 2011;47(3):333–338. doi:10.1016/j.ejca.2010.10.024

8. Ibrahim TR, Abdel-Raouf SM. Immunohistochemical study of Glypican-3 and HepPar-1 in differentiating hepatocellular carcinoma from metastatic carcinomas in FNA of the liver. Pathol Oncol Res. 2015;21(2):379–387. doi:10.1007/s12253-014-9830-6

9. Libbrecht L, Severi T, Cassiman D, et al. Glypican-3 expression distinguishes small hepatocellular carcinomas from cirrhosis, dysplastic nodules, and focal nodular hyperplasia-like nodules. Am J Surg Pathol. 2006;30(11):1405–1411. doi:10.1097/01.pas.0000213323.97294.9a

10. Coston WMP, Loera S, Lau SK, et al. Distinction of hepatocellular carcinoma from benign hepatic mimickers using Glypican-3 and CD34 immunohistochemistry. Am J Surg Pathol. 2008;32(3):433–444. doi:10.1097/PAS.0b013e318158142f

11. Liu S, Wang M, Zheng C, Zhong Q, Shi Y, Han X. Diagnostic value of serum glypican-3 alone and in combination with AFP as an aid in the diagnosis of liver cancer. Clin Biochem. 2020;79:54–60. doi:10.1016/j.clinbiochem.2020.02.009

12. Li B, Liu H, Shang HW, Li P, Li N, Ding HG. Diagnostic value of glypican-3 in alpha fetoprotein negative hepatocellular carcinoma patients. Afr Health Sci. 2013;13(3):703–709. doi:10.4314/ahs.v13i3.26

13. Chen M, Li G, Yan J, et al. Reevaluation of glypican-3 as a serological marker for hepatocellular carcinoma. Clin Chim Acta. 2013;423:105–111. doi:10.1016/j.cca.2013.04.026

14. Nault JC, Guyot E, Laguillier C, et al. Serum proteoglycans as prognostic biomarkers of hepatocellular carcinoma in patients with alcoholic cirrhosis. Cancer Epidemiol Biomark Prev. 2013;22(8):1343–1352. doi:10.1158/1055-9965.EPI-13-0179

15. Zheng X, Liu X, Lei Y, Wang G, Liu M. Glypican-3: a novel and promising target for the treatment of hepatocellular carcinoma. Front Oncol. 2022;12:824208. doi:10.3389/fonc.2022.824208

16. Feng J, Zhu R, Chang C, et al. CK19 and Glypican 3 expression profiling in the prognostic indication for patients with HCC after surgical resection. PLoS One. 2016;11(3):e0151501. doi:10.1371/journal.pone.0151501

17. Dong SY, Sun W, Xu B, et al. Quantitative image features of gadoxetic acid-enhanced MRI for predicting glypican-3 expression of small hepatocellular carcinoma ≤3 cm. Clin Radiol. 2023;78(10):e764–e772. doi:10.1016/j.crad.2023.06.009

18. Chong H, Gong Y, Zhang Y, Dai Y, Sheng R, Zeng M. Radiomics on gadoxetate Disodium-enhanced MRI: non-invasively identifying glypican 3-positive hepatocellular carcinoma and postoperative recurrence. Acad Radiol. 2023;30(1):49–63. doi:10.1016/j.acra.2022.04.006

19. Zhu L, Tang W, Jiang B, et al. Noninvasive prediction of glypican-3 expression and recurrence-free survival in hepatocellular carcinoma using contrast-enhanced ultrasound. Cancer Imaging. 2026;26(1):67. doi:10.1186/s40644-026-01033-9

20. Wang Y, Shen Z, Zhu Z, Han R, Huai M. Clinical values of AFP, GPC3 mRNA in peripheral blood for prediction of hepatocellular carcinoma recurrence following OLT: AFP, GPC3 mRNA for prediction of HCC. Hepat Mon. 2011;11(3):195–199.

21. Chen Y, Qin Y, Wu Y, et al. Preoperative prediction of glypican-3 positive expression in solitary hepatocellular carcinoma on gadoxetate-disodium enhanced magnetic resonance imaging. Front Immunol. 2022;13:973153. doi:10.3389/fimmu.2022.973153

22. Liang X, Li L, Wang Y, et al. US molecular imaging of glypican-3 expression in hepatocellular carcinoma using targeted biosynthetic gas vesicles. Radiol Imaging Cancer. 2026;8(2):e250480. doi:10.1148/rycan.250480

23. Gu D, Xie Y, Wei J, et al. MRI-based radiomics signature: a potential biomarker for identifying glypican 3-positive hepatocellular carcinoma. J Magn Reson Imaging. 2020;52(6):1679–1687. doi:10.1002/jmri.27199

24. Sun S, Xiao S, Jiang Z, et al. Radiomic analysis of contrast-enhanced CT predicts glypican 3-positive hepatocellular carcinoma. Curr Med Imaging. 2024;20:e15734056277475. doi:10.2174/0115734056277475240215115629

25. Wang X, Li L, Wang L, Chen M. The expression of Ki-67 and Glypican −3 in hepatocellular carcinoma was evaluated by comparing DWI and 18F-FDG PET/CT. Front Oncol. 2023;13:1026245. doi:10.3389/fonc.2023.1026245

26. Zhao J, Gao S, Sun W, et al. Magnetic resonance imaging and diffusion-weighted imaging-based histogram analyses in predicting glypican 3-positive hepatocellular carcinoma. Eur J Radiol. 2021;139:109732. doi:10.1016/j.ejrad.2021.109732

27. Gao Y, Liu D, Miao Y, et al. Radiogenomic MRI biomarkers for noninvasive prediction of GPC3 expression and tumor microenvironment in hepatocellular carcinoma. J Transl Med. 2025. doi:10.1186/s12967-025-07504-0

28. Zhang J, Zhu X, Qiu J, et al. Noninvasive prediction of Glypican-3 expression in hepatocellular carcinoma using Habitat-based and peritumoral CT radiomics: a nomogram approach. Cancer Imaging. 2025;25(1):144. doi:10.1186/s40644-025-00966-x

29. Zhao C, Zhou Z, Zhang J, et al. A novel dynamic nomogram based on contrast-enhanced computed tomography radiomics for prediction of glypican-3-positive hepatocellular carcinoma. Front Oncol. 2025;15:1640697. doi:10.3389/fonc.2025.1640697

30. Li N, Dong T, Wang P, Li Q, Nie F. Predicting glypican-3 expression in hepatocellular carcinoma: a comprehensive analysis using combined contrast-enhanced ultrasound and clinical factors. Clin Hemorheol Microcirc. 2023;85(4):407–420. doi:10.3233/CH-231912

31. Huang F, Pang J, Wu Y, et al. Development of a model combining CEUS LI-RADS and clinical features for predicting glypican-3 expression in hepatocellular carcinoma. Abdom Radiol (NY). 2025. doi:10.1007/s00261-025-04861-8

32. Shu Q, Pang C, Wang Z, et al. Contrast-enhanced ultrasound and clinical features for preoperative prediction of GPC-3 expression in hepatocellular carcinoma. Abdom Radiol (NY). 2026. doi:10.1007/s00261-026-05388-2

33. Zhang Y, Liu R, Li P, et al. Preoperative prediction of Glypican-3 in hepatocellular carcinoma by two-dimensional ultrasound, contrast-enhanced ultrasound and microflow imaging. Clin Hemorheol Microcirc. 2025;91(1):31–38. doi:10.1177/13860291251376273

34. Li L, Zou X, Zheng W, et al. Contrast-enhanced US with sulfur hexafluoride and perfluorobutane: LI-RADS for diagnosing hepatocellular carcinoma. Radiology. 2023;308(2):e230150. doi:10.1148/radiol.230150

35. Morel DR, Schwieger I, Hohn L, et al. Human pharmacokinetics and safety evaluation of SonoVue, a new contrast agent for ultrasound imaging. Invest Radiol. 2000;35(1):80–85. doi:10.1097/00004424-200001000-00009

36. Yanagisawa K, Moriyasu F, Miyahara T, Yuki M, Iijima H. Phagocytosis of ultrasound contrast agent microbubbles by Kupffer cells. Ultrasound Med Biol. 2007;33(2):318–325. doi:10.1016/j.ultrasmedbio.2006.08.008

37. Watanabe R, Matsumura M, Munemasa T, Fujimaki M, Suematsu M. Mechanism of hepatic parenchyma-specific contrast of microbubble-based contrast agent for ultrasonography: microscopic studies in rat liver. Invest Radiol. 2007;42(9):643–651. doi:10.1097/RLI.0b013e31805f2682

38. Takai H, Ashihara M, Ishiguro T, et al. Involvement of glypican-3 in the recruitment of M2-polarized tumor-associated macrophages in hepatocellular carcinoma. Cancer Biol Ther. 2009;8(24):2329–2338. doi:10.4161/cbt.8.24.9985

39. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749–762. doi:10.1038/nrclinonc.2017.141

40. Liang L, Pang JS, Gao RZ, et al. Development and validation of a combined radiomic and clinical model based on contrast-enhanced ultrasound for preoperative prediction of CK19-positive hepatocellular carcinoma. Abdom Radiol (NY). 2025;50(8):3516–3529. doi:10.1007/s00261-025-04799-x

41. Zhang D, Zhang XY, Lu WW, et al. Predicting Ki-67 expression in hepatocellular carcinoma: nomogram based on clinical factors and contrast-enhanced ultrasound radiomics signatures. Abdom Radiol (NY). 2024;49(5):1419–1431. doi:10.1007/s00261-024-04191-1

42. Zheng R, Zhang X, Liu B, et al. Comparison of non-radiomics imaging features and radiomics models based on contrast-enhanced ultrasound and Gd-EOB-DTPA-enhanced MRI for predicting microvascular invasion in hepatocellular carcinoma within 5 cm. Eur Radiol. 2023;33(9):6462–6472. doi:10.1007/s00330-023-09789-5

43. Chen K, Zhang S, Huang T, et al. CEUS-based microvascular invasion predictor in HCC: improving prognostic stratification following thermal ablation. Liver Int. 2026;46(6):e70661. doi:10.1111/liv.70661

44. Yang D, Wu W, Zhang Q, et al. Quantitative sonazoid contrast-enhanced ultrasound features for predicting microvascular invasion of hepatocellular carcinoma. Eur J Radiol. 2026;194:112477. doi:10.1016/j.ejrad.2025.112477

45. Kim TK, Noh SY, Wilson SR, et al. Contrast-enhanced ultrasound (CEUS) liver imaging reporting and data system (LI-RADS) 2017 - a review of important differences compared to the CT/MRI system. Clin Mol Hepatol. 2017;23(4):280–289. doi:10.3350/cmh.2017.0037

46. Fu SJ, Qi CY, Xiao WK, Li SQ, Peng BG, Liang LJ. Glypican-3 is a potential prognostic biomarker for hepatocellular carcinoma after curative resection. Surgery. 2013;154(3):536–544. doi:10.1016/j.surg.2013.02.014

47. Russell BC, Torralba A, Murphy KP, Freeman WT. LabelMe: a database and web-based tool for image annotation. Int J Comput Vis. 2008;77(1):157–173. doi:10.1007/s11263-007-0090-8

48. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594. doi:10.1136/bmj.g7594

49. Li C, Liu Y, Wu M, Dai W, Song J, Wang Y. Kupffer phase radiomics signature in sonazoid contrast-enhanced ultrasound predicts immunohistochemistry marker expression in hepatocellular carcinoma. Cancer Med. 2025;14(19):e71153. doi:10.1002/cam4.71153

50. Morford LA, Davis C, Jin L, Dobierzewska A, Peterson ML, Spear BT. The oncofetal gene glypican 3 is regulated in the postnatal liver by zinc fingers and homeoboxes 2 and in the regenerating liver by alpha-fetoprotein regulator 2. Hepatology. 2007;46(5):1541–1547. doi:10.1002/hep.21825