Noninvasive prediction of Glypican-3 expression in hepatocellular carcinoma using Habitat-based and peritumoral CT radiomics: a nomogram approach.

Background
Hepatocellular carcinoma (HCC) ranks as the sixth most common malignancy and the third leading cause of cancer-related mortality worldwide [1]. Glypican-3 (GPC3) plays a critical role in regulating cellular growth and differentiation, establishing itself as a key biomarker for targeted therapy and prognostic evaluation in HCC [2]. Current clinical practice relies on immunohistochemical assessment of GPC3 expression; however, this biopsy-dependent approach carries inherent risks of invasive procedures; furthermore, it is fundamentally limited by sampling bias that fails to capture tumor heterogeneity [3, 4].
Radiomics has demonstrated potential for noninvasive quantitative analysis of HCC in recent years. However, radiomics models primarily reflect macroscopic tumor characteristics and fail to account for intratumoral phenotypic heterogeneity [5]. Peritumoral radiomics models address this limitation by extracting imaging features from expansion regions beyond the tumor boundary (typically defined as 0.3–3 cm from the tumor margin) [6]. These models, when integrated with machine learning algorithms, enhance the predictive accuracy of prognostic biomarkers [7, 8]. Habitat imaging further advances this paradigm by partitioning intratumoral biological subregions through cluster analysis of voxels with similar textural parameters; this approach quantifies textural features and spatial topological relationships, thereby enabling comprehensive characterization of intratumoral spatial heterogeneity [9, 10].
This study developed a predictive framework integrating habitat imaging for intratumoral heterogeneity analysis with peritumoral Computed Tomography (CT) radiomics, establishing a composite model for noninvasive assessment of GPC3 expression in HCC. This approach delivers quantifiable and interpretable imaging biomarkers to inform personalized therapeutic decision-making for HCC patients.

Methods

Patients
This retrospective study was approved by our institutional ethics committee with waiver of informed consent (Approval No. QT2025045). Among 374 patients with surgically confirmed HCC treated between January 2021 and January 2024, inclusion criteria required: (1) age ≥ 18 years; (2) Child-Pugh class A or B; (3) histologically confirmed HCC with documented GPC3 status, microvascular invasion (MVI), and Edmondson grade; (4) preoperative contrast-enhanced CT within 1 week before surgery; (5) no prior antitumor therapy; (6) selection of the dominant lesion correlating with pathological diagnosis in multifocal cases. Exclusion criteria encompassed: (1) suboptimal image quality (severe artifacts/incomplete sequences; n = 18); (2) incomplete clinical/pathological data (n = 59); (3) current/past malignancies (n = 94). After excluding 171 ineligible patients, 203 subjects were enrolled and randomly divided into training (n = 142) and validation (n = 61) sets at 7:3 ratio for model development and testing (Fig. 1).

CT examination protocol
All examinations were performed using either a Siemens SOMATOM Definition AS 128 or Siemens SOMATOM Definition AS 40 CT scanner. All patients underwent non-contrast abdominal CT followed by contrast-enhanced scans during the arterial phase (AP) and portal venous phase (VP). (Supplementary data S1)

Baseline evaluation
Demographic characteristics, tumor features, and clinicopathological data were collected, including: age, gender, Edmondson grade, MVI status, Ki-67, CD10, CD34, hepatitis B virus (HBV) infection status, alpha-fetoprotein (AFP, dichotomized at 20 µg/L), total bilirubin (TBIL), albumin (ALB), prothrombin time (PT).

Radiological image analysis
The assessed imaging characteristics included tumor size, tumor capsule integrity, margin delineation, intratumoral necrosis, and peritumoral manifestations. All CT images were initially interpreted by an abdominal radiologist with 8 years of experience, followed by independent verification and classification by a second abdominal radiologist with 15 years of experience. In cases of interobserver discrepancy, a third radiologist with 24 years of experience was consulted to reach consensus. All evaluations were performed blinded to clinical and pathological information to ensure objective and independent assessments.

Pathological examination
Two pathologists with ≥ 8 years of diagnostic experience independently assessed GPC3 expression using the Immunoreactive Score (IRS) system proposed by Remmele and Stegner. The IRS was calculated by multiplying the staining intensity score (0 = no staining; 1 = weak; 2 = moderate; 3 = strong) by the percentage of positive cells (0 = < 1%; 1 = 1%-10%; 2 = 11%-50%; 3 = 51%-80%; 4 = 81%-100%), yielding a total score ranging from 0 to 12 [11]. An IRS ≥ 2 was defined as a positive expression, while an IRS < 2 was considered negative [12].

Tumor segmentation and extraction of peritumor radiomics feature
Preoperative VP CT images in DICOM format were processed using the open-source software ITK-SNAP (version 4.0; http://www.itksnap.org). Volumes of interest (VOI) were manually delineated slice-by-slice along the entire tumor contour by an abdominal radiologist with 8 years of experience. To assess interobserver reproducibility, a second abdominal radiologist with 15 years of experience independently segmented 30 randomly selected cases. Feature extraction consistency was quantified using the intraclass correlation coefficient (ICC), with ICC values >0.80 indicating excellent agreement. In cases of multifocal disease, only the largest lesion was selected for analysis. Subsequent radiomics feature extraction was performed using PyRadiomics (version 3.0.1) [13], including: first-order statistics (e.g., intensity histogram features), second-order texture features (e.g., gray-level co-occurrence matrix, gray-level run-length matrix), higher-order features (e.g., wavelet-transformed morphological parameters). A total of 1315-dimensional raw features were extracted from each 5 mm and 8 mm peritumoral regions.

Habitat generation
Local features including entropy and energy values were extracted from each voxel within volumes of interest (VOIs). These feature vectors represented distinct aspects of voxel properties. The Calinski-Harabasz (CH) index method determined an optimal cluster number of 3 for habitat partitioning. Subsequently, K-means clustering was performed to classify subregions within each sample. Ultimately, 1315 radiomics features were calculated per habitat, yielding 3947 features across three habitats (n = 1315 × 3).

Feature selection and model construction
In this retrospective cohort of 203 patients, radiomics features underwent comprehensive preprocessing. Continuous variables were scaled to the range [0, 1] using min-max normalization. Outliers were addressed through median imputation. For clinical variables, dimensionality reduction was implemented via backward stepwise regression to enhance model interpretability and mitigate overfitting risks.
Peritumoral radiomics feature selection was performed using random forests (RF). For both 5 mm and 8 mm peritumoral expansions (each initially comprising 1315 features), the maximum relevance minimum redundancy (mRMR) algorithm preselected 20 features per expansion to mitigate overfitting. Through recursive feature elimination with random forests, the top 3 decision trees from 5 mm features and 5 trees from 8 mm features were identified as significant model contributors [14]. The cohort was partitioned into training (n = 142) and validation (n = 61) sets at 7:3 ratio. SHapley Additive exPlanations (SHAP) analysis quantified feature importance and directional impact on prediction probability, enhancing model interpretability [15]. The final radiomics Rad-score derived from this process served as the peritumoral risk metric.
The habitat analysis modeling employed the minimum Redundancy Maximum Relevance (mRMR) algorithm to select the top 20 features for each modality. Random forest was then applied to calculate importance scores for these 20 features. Following importance-based ranking in descending order, redundant features were eliminated, ultimately retaining the top 3 most important features for habitat model construction. The final Risk Score output served as the habitat model’s risk metric. The complete workflow is illustrated in Fig. 2.

The peritumoral Rad-score and habitat Risk Score were integrated with clinically significant variables selected via logistic regression to develop a predictive nomogram. This hybrid architecture combined imaging feature representations with clinical parameters. Interdependencies among clinical indicators, peritumoral features, and habitat analysis were visualized using chord diagrams to quantify feature interactions.

Statistical analysis
All statistical analyses were conducted using R software (version 4.4.2; http://www.Rproject.org). Continuous variables were compared using Student’s t test or Mann-Whitney U test as appropriate, while categorical variables were analyzed with χ² test or Fisher exact test. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with calculation of the area under the curve (AUC), accuracy, sensitivity, and specificity for each ROC curve. Decision curve analysis (DCA) was performed to assess clinical utility across various threshold probabilities, and the Delong test was used to compare AUC values between different models. By amalgamating clinical, peritumoral and habitat features, a nomogram was formulated. Chord diagrams were generated to visualize correlations among clinical parameters, radiomics features, and habitat characteristics. A two-sided p value less than 0.05 was considered statistically significant.

Results

Patient characteristics
The cohort of 203 patients was stratified into training (n = 142) and validation (n = 61) sets. Detailed baseline characteristics for both cohorts are summarized in Table 1. Comparative analysis demonstrated no significant differences in clinical or radiologic features between training and validation groups (all p > 0.05), ensuring balanced cohort comparability.

Establishment of clinical models
Seventeen clinical features underwent rigorous selection using logistic regression analysis, from which five demonstrated significant predictive value (Fig. 3): Edmondson grade, MVI, Ki-67 proliferation index, HBV infection status, and PT. These selected features were incorporated into a logistic regression model to generate a clinical risk score, establishing the final clinical prediction model.

Model construction for the prediction of GPC3
Eight predictive models were established: a clinical feature model, a habitat feature model, a 5 mm peritumoral (EX5MM) radiomics model, an 8 mm peritumoral (EX8MM) radiomics model, a combined Clinics_Habitat model, a Clinics_EX5MM model, a Clinics_EX8MM model, and an integrated GPC3-RadNomogram model.

Model evaluation and visualization
ROC curve analysis demonstrated that the Habitat model alone achieved AUCs of 0.822 (95% CI: 0.755–0.889) and 0.792 (95% CI: 0.681–0.903), the integrated GPC3-RadNomogram model showed superior predictive performance compared to the Clinics and the Habitat models, with the AUC values of 0.912 (95% CI: 0.866–0.958) in the training cohort and 0.927 (95% CI: 0.861–0.993) in the validation cohort (Fig. 4A, B), respectively. These findings supported the development of a clinically applicable nomogram (Fig. 5). Calibration curves revealed excellent agreement between predicted and observed outcomes (Hosmer-Lemeshow test: p = 0.93 for training cohort; p = 0.67 for validation cohort) (Fig. 4C, D). Decision curve analysis confirmed greater clinical net benefit across threshold probabilities (15% − 60%) for the integrated GPC3-RadNomogram model versus individual models approaches (Fig. 4E, F). The integrated model demonstrated statistically significant improvement in discriminatory capacity compared to individual models (DeLong test: all |z|>1.96, p < 0.001; Tables S4, S5). The chord diagram (Fig. 6) effectively delineates interfeature relationships within the GPC3-RadNomogram, demonstrating low multicollinearity through predominantly thin ribbons while highlighting clinically relevant associations via select thicker connections that substantiate our model’s integration of clinical, intratumoral heterogeneity, and peritumoral characteristics.

Discussion
GPC3, encoded by the RAMP3 gene (one of five key prognostic genetic factors), demonstrates upregulated expression associated with enhanced tumor aggressiveness and activation of pro-angiogenic pathways in HCC [16]. As a membrane-anchored coreceptor, GPC3 drives HCC malignant progression and metastasis by coordinating three oncogenic mechanisms: potentiation of proliferative signaling, suppression of programmed cell death, and remodeling of the extracellular microenvironment [17]. Crucially, GPC3 overexpression occurs in >70% of HCC cases, while being nearly undetectable in normal liver tissue, cirrhotic nodules, and benign hepatic lesions. This discriminative capacity significantly outperforms conventional serum tumor biomarkers [18]. Consequently, noninvasive prediction of GPC3 expression status holds substantial clinical value for guiding targeted therapies (e.g., anti-angiogenic agents) and prognostic stratification.
This study integrates peritumoral radiomics features with intratumoral habitat analysis, simultaneously capturing microenvironmental alterations in peritumoral regions and intratumoral biological heterogeneity, thereby overcoming the conventional radiomics limitation of treating tumors as homogeneous entities. K-means clustering was applied for habitat segmentation to quantitatively characterize subregions with distinct biological behaviors within tumors [19]. Building upon this approach, we developed a multimodal model for predicting GPC3 expression in HCC. The results demonstrate that this integrated model - incorporating intratumoral heterogeneity (habitat), peritumoral features, and clinical factors - significantly outperforms individual model prediction models relying solely on either peritumoral radiomics or clinical parameters alone.
The DeLong test confirmed statistically significant improvements in discriminatory capacity for our integrated GPC3-RadNomogram model compared to individual models (clinical: z = -2.57, p < 0.05; EX5MM: z = -2.97, p < 0.01; EX8MM: z = -3.39, p < 0.001). Our model provides clinically actionable guidance for treatment personalization in hepatocellular carcinoma based on predicted GPC3 expression levels. For patients with high predicted probability of GPC3 overexpression (>80%), first-line consideration should be given to GPC3-targeted immunotherapies including CAR-T cell therapy or antibody-drug conjugates to maximize therapeutic efficacy [20, 21]. In contrast, conventional treatment regimens may be more appropriate for patients with low predicted GPC3 expression (< 30%) to minimize unnecessary overtreatment. Decision curve analysis substantiates the clinical utility of our GPC3-RadNomogram integrated model, demonstrating significant net benefit across the 15%-40% threshold probability range. For this critical patient subgroup, we advocate a comprehensive monitoring protocol consisting of bimonthly GPC3 biomarker assessment complemented by functional imaging evaluation to enable real-time therapeutic response monitoring and timely treatment adjustment. These findings establish that combining intratumoral habitat analysis with peritumoral imaging features enables noninvasive prediction of GPC3 expression status in HCC. This approach delivers quantifiable and interpretable imaging biomarkers to inform personalized therapeutic decision-making for HCC patients.
Prior studies have established that incorporating peritumoral radiomics features enhances predictive performance for HCC biomarkers and supports clinical decision-making [22–24]. The peritumoral 2–8 mm region has been shown to contain critical biological information regarding MVI and HBV infection, with integrated models combining intratumoral features and 2 mm peritumoral regions demonstrating optimal performance (mean internal validation AUC = 0.831; maximum external validation AUC = 0.839) [25]. These findings support our strategy of combining peritumoral radiomics with intratumoral habitat analysis. Hongjie Hu et al. [26] further demonstrated that radiomics models (Rad-score) based on 10 mm peritumoral features significantly outperformed clinical nomograms in predicting recurrence-free survival (AUC = 0.809 to 0.892 versus clinical models: 0.670 to 0.680). Our results align with these conclusions, showing consistent superiority of peritumoral models over clinical models in both training (EX5MM AUC = 0.777, 95% CI: 0.699–0.855; EX8MM AUC = 0.773, 95% CI: 0.694–0.852) and validation cohorts (EX5MM AUC = 0.747, 95% CI: 0.626–0.869; EX8MM AUC = 0.741, 95% CI: 0.618–0.865 versus clinical model: training AUC = 0.71, 95% CI: 0.624–0.795; validation AUC = 0.781, 95% CI: 0.661–0.902), underscoring the prognostic value of peritumoral radiomics in HCC. Our study further confirms the necessity of integrating radiomics with clinical indicators, consistent with Cuiyun Wu et al. [27], where combining radiomics features with clinical parameters (e.g., AFP level, Edmondson grade) significantly improved model performance (training AUC = 0.884; validation AUC = 0.819).
Building upon previous methodology, Jingran Wu et al. [28] developed five distinct predictive models by integrating intratumoral, peritumoral, and habitat radiomics features from CT images. Their results demonstrated the superior performance of habitat-derived radiomics features in discriminating EGFR mutation status in stage I non-small cell lung cancer (NSCLC), significantly outperforming conventional peritumoral models across training (AUC = 0.886), validation (AUC = 0.812), and external test cohorts (AUC = 0.790). Similarly, Zhao et al. [29] employed PET/CT-based habitat subregion segmentation to precisely quantify colorectal cancer heterogeneity, with the habitat model demonstrating robust predictive capacity for KRAS/NRAS/BRAF mutations (training AUC = 0.759; validation AUC = 0.701). Our habitat-based analysis provides novel insights into the spatially heterogeneous expression of Glypican-3 (GPC3) in HCC. The highest weighted feature Habitat1_wavelet_HLH_firstorder_Skewness, derived from three-dimensional HLH wavelet transformation within Habitat 1 subregions, this habitat-specific radiomic feature enables spatial characterization of heterogeneity in tumor microenvironment through quantifying the spatial heterogeneity of cellular density distribution and distribution patterns of potential resistance foci, which captures GPC3-associated biological processes at a resolution unachievable by conventional biopsy techniques [30]. GPC3-positive hepatocellular carcinomas typically demonstrate more aggressive biological behavior, characterized by heterogeneous cellular distribution, diversified vascular supply, and distinct subregional proliferation patterns - features that are precisely quantified by our habitat analysis [31]. This finding aligns with established pathological mechanisms wherein GPC3 overexpression drives tumor heterogeneity through activation of the Wnt signaling pathway, generating microscopic tumor subregions that conventional radiomics approaches fail to characterize [32]. These findings, consistent with our current study, collectively establish habitat analysis as a noninvasive tool for quantifying intratumoral heterogeneity. When synergistically combined with radiomics and imaging biomarkers, this approach provides novel pathways for precision treatment decision-making in clinical oncology practice.
This retrospective single-center study has inherent methodological limitations. The homogeneous patient population and standardized imaging protocols from a single institution may introduce selection bias and limit the model’s generalizability to broader clinical settings. While the cohort of 203 patients provides an adequate sample size for preliminary model development, it remains insufficient for implementing more sophisticated algorithms like deep learning and carries inherent overfitting risks. Future investigations should prioritize multicenter prospective validation to establish model robustness across different institutions and scanner platforms, explore the integration of multimodal imaging to fully characterize tumor hemodynamics, and systematically determine optimal peritumoral boundary thresholds to enhance predictive accuracy. This study primarily focused on evaluating the diagnostic performance of GPC3 as a biomarker, with its association to clinical endpoints such as patient survival outcomes not yet validated. In subsequent investigations, we plan to systematically assess the clinical value of this radiomic model for guiding personalized treatment strategies and prognostic stratification in hepatocellular carcinoma using continuously updated long-term follow-up data. These refinements will be critical for clinical translation and implementation.

Conclusion
This study developed a novel CT habitat radiomics model that significantly improves preoperative prediction of GPC-3 expression in HCC through integrated analysis of intratumoral spatial heterogeneity, peritumoral microenvironment features, and clinicopathological parameters. The validated GPC3-RadNomogram composite nomogram demonstrates robust diagnostic performance, providing a noninvasive alternative to biopsy-dependent biomarker evaluation. Habitat-driven quantification outperforms conventional radiomics approaches, confirming its critical role in advancing precision medicine for HCC management.

Supplementary Information
Below is the link to the electronic supplementary material.