Multiparametric dual-energy computed tomography radiomics for predicting microvascular invasion in hepatocellular carcinoma.

Introduction
Hepatocellular carcinoma (HCC) accounts for 75%–85% of primary liver cancers [1]. Despite curative treatments like liver resection and transplantation, HCC is associated with a poor prognosis and high recurrence rates, with a 70% recurrence rate within 5 years post-surgery and 35% after transplantation [2]. Tumor stage and histopathological indicators of biological aggressiveness are the key prognostic factors influencing management and treatment decisions in HCC. Among these characteristics, microvascular invasion (MVI) is recognized as a crucial predictor of poor prognosis [3]. MVI leads to increased malignancy of tumor cells and destruction of surrounding tissue architecture, indicating a higher likelihood of tumor metastasis [4]. Studies indicate that in resected HCC, patients with MVI have a 5-year survival rate of 26–67%, compared to 53–83% in those without MVI [5]. At present, the diagnosis of MVI in clinical practice can only be confirmed through histopathological examination of surgical specimens [6]. Therefore, accurate preoperative assessment of MVI is crucial for treatment planning and prognosis of patients, such as guiding wide margin resections and selecting suitable candidates for liver transplantation [7, 8].
Computed tomography (CT) plays a vital role in the clinical management of HCC. Several qualitative imaging features, including non-smooth tumor margins, irregular peritumoral arterial enhancement, and internal arteries, are established risk factors for MVI in HCC [9, 10]. However, their diagnostic performance is hampered by relatively low interobserver agreement and sensitivity [11]. Consequently, there has been an increasing focus on quantitative imaging biomarkers to overcome these limitations.
Dual-energy CT (DECT) acquires data at two distinct X-ray energy levels, generating virtual monochromatic images (VMIs) comparable to conventional CT and providing material-specific density maps, such as iodine density (ID) and effective atomic number (Zeff) maps [12]. Conventional quantitative parameters derived from DECT, such as iodine concentration (IC) and Zeff, have been applied in the preoperative diagnosis of MVI in HCC [13–16]. However, the intrinsic heterogeneity of HCC limits the ability of these parameters to fully characterize tumor biology, resulting in suboptimal diagnostic performance of current DECT-based approaches [17].
Radiomics is an emerging quantitative technique that extracts and analyzes high-throughput features from medical images, enabling the quantitative assessment of tumor heterogeneity [18]. This methodology has demonstrated significant predictive value across multiple clinical domains in HCC, including MVI assessment [2, 6, 19, 20], molecular subtype classification [21], immune status prediction [22], histological grading [23], and treatment response evaluation [24, 25]. Nevertheless, it remains unclear whether multiparametric DECT-based radiomics, providing additional quantitative information beyond conventional CT, can enhance the noninvasive diagnosis of MVI in HCC.
This study aimed to combine multiparametric DECT maps with radiomics for the preoperative prediction of MVI in HCC. It also compares the performance with quantitative parameters. Figure 1 shows the overall framework of the proposed methodology.

Materials and methods

Study population
This retrospective study was approved by the Ethics Committee of Sun Yat-sen Memorial Hospital, Sun Yat-sen University, and written informed consent was waived due to its retrospective design. The dataset comprised patients with pathologically confirmed HCC who underwent contrast-enhanced DECT scans at Sun Yat-sen Memorial Hospital, Sun Yat-sen University, between January 1, 2021, and June 20, 2024 (see Fig. 2). The inclusion criteria were: (1) pathologically confirmed primary, solitary HCC; (2) DECT examination performed within one month before surgery; and (3) no history of prior HCC-related treatments (e.g., radiotherapy, chemotherapy, radiofrequency ablation). Exclusion criteria included: (1) evidence of macroscopically distinguishable portal vein branch tumor thrombus on imaging; (2) nondiagnostic image quality (e.g., significant motion artifacts); and (3) incomplete imaging or clinicopathological data. A detailed description of the inclusion and exclusion criteria is provided in Supplementary S1. Eligible patients were randomly assigned to training and test cohorts at a 7:3 ratio. Clinical characteristics were retrieved from the hospital information system.

Imaging acquisition and preprocessing
Liver imaging was performed using a Discovery CT750 HD scanner (GE Healthcare) for both non-contrast (NCP) and contrast-enhanced spectral data acquisition. Contrast-enhanced scanning commenced after intravenous injection of iopromide (Ultravist 370, Bayer). The arterial phase (AP) was triggered by automated bolus tracking with a threshold of 100 Hounsfield Units (HU) in the abdominal aorta, while the portal venous (PVP) and delayed (DP) phases acquired at 50–70 and 180 s, respectively. Further scanning protocol details are provided in Supplementary Table S1. Multiparametric maps for each phase, comprising virtual monochromatic images (VMIs) at 70 keV, iodine density (ID) maps, and effective atomic number (Zeff) maps, were reconstructed using an AW80 Workstation (GE Healthcare) for subsequent radiomics analysis. The 70 keV energy level was selected for VMIs because it provides soft-tissue attenuation comparable to that of conventional 120 kVp CT in abdominal imaging [26], serving as a surrogate for conventional CT for comparative analysis. ID maps were generated by material decomposition using an iodine–water basis pair.
Windowing techniques defined the grayscale ranges for these parametric maps, removing information irrelevant to diagnosis and aligning the images with radiologists’ visual perception. Specifically, the window width and level were set to 350/50 HU for VMIs and 15/5 mg/mL for ID maps, and 8/10 (dimensionless) for Zeff maps. These settings were determined by consensus between two radiologists with 5 and 15 years of diagnostic experience, respectively, leveraging their extensive clinical expertise.

Lesion segmentation
Lesion segmentation was performed on the PVP using 3D Slicer (v5.1.0) by two radiologists, R1 with 5 years and R2 with 15 years of abdominal imaging experience, both blinded to clinical and pathological data. R1 first selected the parametric map (e.g., VMI) with the clearest tumor margin in this phase for manual region of interest (ROI) delineation, and the resulting mask was then copied to the other parametric maps (e.g., ID and Zeff). The parametric maps of the remaining phases (NCP, AP, and DP) were registered to the PVP, and the corresponding masks were subsequently propagated. Finally, R2 carefully reviewed all masks and performed manual refinements when necessary.

Radiomics feature extraction and quantitative parameter measurement
For radiomics analysis, all parametric maps were resampled to an isotropic voxel size of 1 mm³, and radiomics features were extracted using PyRadiomics (v3.1.0) [27] from both the original and transformed images, yielding 2120 features per parametric map. The detailed PyRadiomics settings are summarized in Supplementary Table S2. Subsequently, four feature sets were constructed: VMI, ID, Zeff, and Combined MIZ. The VMI, ID, and Zeff sets contained features extracted from their respective parametric maps across all four phases (NCP, AP, PVP, and DP), whereas the Combined MIZ set integrated all features from the 12 parametric maps.
Additionally, six quantitative parameters—IC, normalized iodine concentration (nIC), CT values at 40 keV and 70 keV (CT40keV, CT70keV), the slope of the spectral HU curve (λHU), and the Zeff value—were measured for each phase, as described in previous studies [13, 15, 16]. Further details can be found in Supplementary S2.

Feature selection
Radiomics features often exhibit redundancy, and high-dimensional data can lead to overfitting and increased computational burden [28]. To address this issue, a three-step feature selection procedure was performed on the training cohort and applied to the test cohort. First, Z-score normalization was used to standardize all features. Second, analysis of variance (ANOVA) was applied to identify features showing significant variability (p < 0.05). Third, to balance information retention and overfitting risk, the minimum redundancy maximum relevance (mRMR) was used to select 50 relevant and low-redundant features (see Supplementary S3 for details). Finally, the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was employed to optimize the regularization parameter λ and select the most predictive features.
Univariate logistic regression was applied to all quantitative parameters to identify predictive variables, and those with a significant association (p < 0.05) were retained for subsequent analysis.

Model construction and evaluation
Model construction was performed using the random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) classifiers for each feature set, resulting in 15 models (see Supplementary Table S3 for parameter settings). All models were trained on the training cohort and validated on the test cohort, with 10-fold cross-validation used to evaluate model performance on the training data. To mitigate class imbalance, the synthetic minority oversampling technique (SMOTE) was applied to the training cohort. Models were developed using Python (v3.12.4) with scikit-learn (v1.4.2) for RF and SVM, and XGBoost (v2.1.1) for XGBoost.
Receiver operating characteristic (ROC) curve analysis was performed to evaluate the discriminative ability of the established models, with the area under the curve (AUC) used as the primary performance metric. Based on the optimal cutoff determined by the maximum Youden index, accuracy, sensitivity, specificity, precision, and F1 score were further calculated.

Statistical analysis
Statistical analysis was performed using Python (v3.12.4). A two-tailed p-value < 0.05 was considered statistically significant. For continuous variables, the independent-samples t-test or Mann–Whitney U test was used, and categorical variables were compared using the chi-square (χ²) test or Fisher’s exact test. The Pearson correlation coefficient (r) was used to evaluate the linear relationships among features. The DeLong test was employed to calculate and compare the AUCs of the models. Calibration curves were generated with 1,000 bootstrap resamples to assess the predictive accuracy of the models. The Hosmer–Lemeshow test was applied to evaluate model stability and goodness-of-fit.

Results

Clinical characteristics
A total of 126 patients (mean age, 56.79 ± 12.07 years; 113 men; 47 MVI-positive) were enrolled and randomly assigned to either a training cohort (n = 88, 33 MVI-positive) or a test cohort (n = 38, 14 MVI-positive). No statistically significant differences were observed in clinical characteristics between the training and test cohorts (see Supplementary Table S4; all p > 0.05).

Feature selection
ANOVA identified a total of 934 features in the Combined MIZ set, 282 in the VMI set, 235 in the ID set, and 417 in the Zeff set. mRMR was then applied to reduce each set to 50 features, prioritizing those with the highest relevance and lowest redundancy. Finally, LASSO further refined the sets, retaining 17 features for the Combined MIZ set, 21 for the VMI set, 20 for the ID set, and 10 for the Zeff set (Supplementary Tables S5–S8). Figure 3 shows the LASSO regression results for the Combined MIZ set feature selection, with those for the other sets presented in Supplementary Figs. S1–S3. The Pearson correlation revealed that both the VMI and Zeff sets had a pair of highly correlated features (r > 0.8), while the Combined MIZ set and ID set showed no such highly correlated feature pairs (Supplementary Figs. S4–S7).
Univariate analysis revealed arterial phase IC (odds ratio [OR] = 2.33, p < 0.001) and nIC (OR = 2.42, p < 0.001) as significant predictors (Table 1).

Model performance
The model performance is summarized in Table 2; Fig. 4. Among the five feature sets, the Combined MIZ set achieved the best performance, with mean AUCs of 0.9129 in the training cohort and 0.8928 in the test cohort. Within this set, the XGBoost classifier achieved the highest AUC values, reaching 0.9427 (95% CI, 0.8995–0.9859) in the training cohort and 0.9375 (95% CI, 0.8681–1.000) in the test cohort. The VMI set achieved mean AUCs of 0.8384 in the training cohort and 0.8125 in the test cohort, with the best performance in the test cohort from RF (AUCs of 0.8347 and 0.8542 in the training and test cohorts, respectively). The ID set yielded mean AUCs of 0.8200 in the training cohort and 0.6984 in the test cohort, with XGBoost performing best (AUCs of 0.8738 and 0.7619 in the training and test cohorts). The Zeff set showed the weakest performance among the VMI and ID sets, with mean AUCs of 0.8112 and 0.6845 in the training and test cohorts, respectively, and XGBoost remained the top-performing classifier (AUCs of 0.8309 and 0.7470 in the training and test cohorts).
Notably, the quantitative parameters alone demonstrated limited diagnostic efficacy, with mean AUCs of 0.6741 in the training cohort and 0.6389 in the test cohort. The RF classifier achieved the highest test cohort AUC in this category (0.7262, 95% CI: 0.5626–0.8898), though its performance was lower than that of the Combined MIZ and other feature sets.

Comparative analysis of model performance
Statistical comparisons using DeLong’s test demonstrated that the Combined MIZ set showed significantly higher AUC values compared with the VMI, ID, Zeff, and quantitative parameter sets in both the training and test cohorts (Table 3: all p < 0.05 for all classifiers). Particularly strong significance was observed for quantitative parameters (p < 0.001 for RF and XGBoost in the training cohort). In contrast, no statistically significant differences were found between the VMI, ID, and Zeff feature sets in either cohort (all p > 0.05).

Calibration curve, Hosmer–Lemeshow test
The calibration curves showed that the predictions from the three models in the Combined MIZ set were in close agreement with the observed outcomes in both the training and test cohorts (Fig. 5). The Hosmer–Lemeshow test results were RF (p = 0.15), SVM (p = 0.73), and XGBoost (p = 0.41) in the training cohort, and RF (p = 0.40), SVM (p = 0.51), and XGBoost (p = 0.21) in the test cohort.

Discussion
This study demonstrates that multiparametric DECT radiomics, which integrates VMIs, ID maps, and Zeff maps, significantly improves the noninvasive prediction of MVI in HCC. The radiomics model, utilizing the Combined MIZ feature set, showed improved performance over single-parametric DECT and conventional quantitative parameters in both the training and test cohorts. Among the classifiers, XGBoost achieved the highest performance, with AUCs of 0.9427 in the training cohort and 0.9375 in the test cohort.
In DECT, 70 keV VMIs are generally considered equivalent to conventional 120 kVp CT images, with a higher signal-to-noise ratio at the same radiation dose [26]. ID values indicate the concentration of iodine contrast within tissues, offering insights into perfusion and vascularity, while Zeff characterizes the chemical composition of the tissue (e.g., hydrogen, carbon, and iodine) [15, 29, 30]. In HCC with MVI, increased microvascular density, endothelial cell proliferation, and structurally incomplete microvasculature lead to elevated vascular permeability [15]. These pathological alterations may disrupt normal tissue perfusion and hemodynamics, and corresponding changes in ID and Zeff may serve as imaging indicators of underlying microcirculatory alterations.
In feature selection, Pearson correlation analysis revealed pairs of features with high correlation (r > 0.8) within the VMI and Zeff sets, indicating potential redundancy within the selected feature subsets. This could be attributed to the uniform selection of 50 features during the mRMR process, which might have included interdependent features with overlapping information. Among the quantitative parameters, arterial phase IC (p < 0.001) and nIC (p < 0.001) were identified as significant predictors, a finding also reported in previous studies [13, 15, 16].
Regarding the performance of the radiomics model, the Combined MIZ set achieved superior AUC compared with the VMI, ID, and Zeff sets across all classifiers in both training and test cohorts (all p < 0.05). This advantage may be associated with the pronounced heterogeneity of HCC, both between different patients (inter-tumoral) and within individual tumors (intra-tumoral) [31].Multiparametric DECT maps, such as VMIs, ID maps, and Zeff maps, each convey distinct aspects of tumor biology. However, relying on a single type of parametric map may be insufficient to capture the full complexity of heterogeneity in HCC. Integrating these complementary maps creates a synergistic representation of diverse tumor characteristics, enabling a more comprehensive and accurate assessment of tumor heterogeneity and ultimately improving predictive performance. Although the mean AUCs of the VMI set were higher than those of the ID and Zeff sets, the differences were not statistically significant in either the training or test cohorts (all p > 0.05). In addition to the limited information provided by each map individually, the inherent overlap in their diagnostic content may further explain the lack of significant performance variation among them.
Notably, the multiparametric DECT radiomics approach outperformed conventional DECT quantitative parameters (Combined MIZ vs. Quantitative parameters). Conventional DECT quantitative parameters (e.g., tumor IC and Zeff) are typically derived from the mean value within a ROI, whereas the first-order features (e.g., Mean) in radiomics capture similar information [27]. Thus, DECT-based radiomics analysis could potentially capture similar information to that provided by traditional quantitative parameters. The key advantage of the radiomics approach lies in the incorporation of more discriminative higher-order texture features (e.g., gray-level co-occurrence matrix, gray-level run-length matrix, and gray-level dependence matrix), which can quantify subtle gray-level distribution patterns and spatial relationships that may not be easily identified through visual inspection [32]. These features more effectively characterize pathological variations within the tumor microenvironment, including disordered cellular architecture, necrotic distribution, angiogenic activity, and extracellular matrix deposition [2], which largely account for the superior performance of the radiomics model.
There were indications of potential overfitting in the single-parametric models, particularly when comparing the AUCs between the training and test cohorts. The mean AUCs of the ID and Zeff models in the training cohort were 0.1216 and 0.1267 higher than those in the test cohort, respectively. This phenomenon may be attributed to the limited sample size (n = 126) and the lack of a feature set based on stability in this study. The relatively small sample size may not adequately represent histological subtypes of HCC with low prevalence, which may reduce the diversity of the training data (we did not perform detailed statistics for each subtype). Moreover, radiomics features are known to be sensitive to variations in imaging protocols, reconstruction algorithms, and segmentation methods. Without rigorous feature stability analysis, unstable features may further exacerbate the risk of overfitting [33]. Nevertheless, the Combined MIZ models demonstrated promising calibration and generalizability. This consistent performance underscores the potential clinical utility of radiomics models incorporating multiparametric DECT maps.
Several recent studies have investigated the potential of DECT quantitative parameters for the preoperative prediction of MVI in HCC. Yang et al. [13] and Kim et al. [14] focused on tumor and peritumoral iodine quantitative parameters, showing that nIC could partly distinguish MVI status (AUC = 0.74–0.87). However, their studies were limited by small sample sizes (50 and 36 patients, respectively) and the lack of validation cohorts to assess generalizability. Lewin et al. [10] combined DECT with perfusion CT parameters and reported that DECT-derived arterial density and iodine concentration, along with perfusion parameters, were significantly associated with MVI. Nevertheless, diagnostic performance metrics for MVI were not reported, and the requirement for additional perfusion scans limits clinical applicability. Zhu et al. [15] developed a scoring model for AFP-negative HCC by integrating DECT quantitative parameters with imaging features, achieving an AUC of 0.792 for the arterial-phase Zeff, which increased to 0.929 after adding imaging features. In contrast to Zhu et al. [15], Lv et al. [34] combined ID values with imaging features, reporting AUCs of 0.872 and 0.904. However, both studies were limited to small HCC subgroups and relied on subjective image interpretation, restricting the generalizability of their models. More recently, Li et al. [16] integrated DECT iodine concentration (IC) with AFP levels and applied PCA for dimensionality reduction, achieving an AUC of 0.84. Compared with the aforementioned studies [13–16, 34], even without incorporating additional variables (e.g., imaging features [15, 34] or laboratory indicators [16]), our radiomics method achieved superior performance (training AUC = 0.9427, test AUC = 0.9375).
Similarly, compared with conventional CT-based radiomics models (AUCs: 0.54–0.93 for training and 0.50–0.89 for test) [2, 5, 19, 20, 35–38], our model achieved higher predictive performance. This improvement may be primarily attributed to the superior image quality of VMI (particularly 70 keV VMI, which provides a relatively higher signal-to-noise ratio [26]) and the additional diagnostic information contributed by the ID and Zeff maps. Furthermore, the application of multiple image transformations (eight types in total, e.g., wavelet and Laplacian of Gaussian) enhanced feature diversity compared with previous studies [2, 5, 20, 35]. In addition, three machine learning classifiers (RF, SVM, and XGBoost) were used, each exhibiting favorable predictive performance within the Combined MIZ set. This finding suggests that our model may possess more robustness than approaches employing a single classifier (e.g., logistic regression) [5, 19, 20, 35, 36], whose performance with alternative algorithms remains unvalidated.
Despite promising results, this study has several limitations. First, it was a single-center retrospective study with a relatively small sample size (n = 126) and lacked an independent external validation cohort. Consequently, the training data may not capture the full histopathological diversity of HCC in broader, real-world populations, which could limit the model’s generalizability. Future studies should use larger, multicenter cohorts to improve external validation and assess the model’s applicability across different clinical settings. Second, our models were developed using radiomics features alone, without incorporating clinical indicators (e.g., AFP levels [16]) or semantic imaging features (e.g., non-smooth tumor margins or irregular peritumoral arterial enhancement [9]). Integrating these additional factors could improve predictive accuracy. Future research should integrate multimodal data to enhance the robustness of prediction models. Third, variations in scanning protocols and patient-specific factors (e.g., lower signal-to-noise ratios in obese patients) may introduce inconsistencies in image quality, which could affect the stability and reproducibility of the radiomics features [33]. Future studies should perform robustness analyses (e.g., intraclass correlation coefficient) to ensure consistent performance across different settings. Finally, although strict inclusion criteria were applied, the exclusion of cases with suboptimal image quality may have introduced selection bias. Future studies should employ objective image quality metrics or advanced reconstruction techniques to minimize such bias.
In conclusion, the multiparametric DECT radiomics model demonstrated promising predictive performance and calibration. These findings suggest that it may serve as a useful noninvasive tool for evaluating MVI status in HCC.

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