Development and validation of a machine learning-based radiomics model using 45-keV virtual monoenergetic images for differentiating fat-poor angiomyolipoma from clear cell renal cell carcinoma

Introduction

Kidney cancer is a common malignancy with a substantial global health burden. In the United States, the American Cancer Society estimated that 80,980 new cases of kidney and renal pelvis cancer would be diagnosed in 2025 (1), while recent epidemiological studies suggest that the global burden of kidney cancer will continue to rise substantially by 2050 (2). Among renal cell carcinoma (RCC) subtypes, clear cell renal cell carcinoma (ccRCC) is the predominant histological subtype in adults. Accurate discrimination between benign and malignant renal masses is critical because the treatment strategies differ substantially. For fat-poor angiomyolipoma (fp-AML), current guidelines recommend active surveillance, reserving intervention for symptomatic or rapidly progressive cases (3,4), whereas small RCCs are often managed with surgical excision, frequently via nephron-sparing approaches (5). Contemporary surgical data from the United States reveal that over 30% of partial nephrectomies yield benign pathology, among which fp-AMLs account for approximately 20% of these benign lesions (6). Therefore, improving preoperative characterization of small renal masses is not only a diagnostic issue but also a means of reducing overtreatment and preserving renal function.

The radiological differentiation between fp-AML and ccRCC remains challenging because of several overlapping imaging features. First, computed tomography (CT) can reliably diagnose classic AML via macroscopic fat identification, but approximately 5% of fp-AMLs lack radiologically demonstrable fat (7), thereby eliminating this hallmark sign. Second, 4–5% of ccRCCs demonstrate lipomatous metaplasia, producing false-positive fat signals that may misleadingly suggest AML. Third, both entities may exhibit overlapping enhancement kinetics (8,9), particularly when fp-AML shows marked corticomedullary hyperenhancement followed by washout, thereby mimicking the typical enhancement pattern of ccRCC. Collectively, these overlapping imaging findings often prompt biopsy or surgery in the absence of definitive benign markers, thereby contributing to unnecessary intervention. This problem is particularly relevant in small renal masses (≤4 cm), which are increasingly detected incidentally yet remain difficult to characterize non-invasively.

At present, differentiation of fp-AML from ccRCC mainly relies on multimodal imaging assessment, including conventional CT, magnetic resonance imaging (MRI), and, more recently, quantitative image analysis. MRI findings such as T2 hypointensity, signal loss on chemical-shift imaging, and diffusion-related characteristics may be helpful for suggesting fp-AML (10); however, substantial overlap persists, particularly in small hypervascular renal masses (11). Likewise, although enhancement-based CT findings are widely used in routine practice, their specificity (SPE) remains limited. Previous quantitative imaging studies, including radiomics-based models, have reported promising results for this differential diagnosis; however, many were derived from conventional CT images, limited feature sets, or single-phase imaging, which may restrict their generalizability (12). Taken together, current imaging methods provide useful clues but still fall short of a consistently reliable non-invasive solution. Therefore, whether spectral CT-derived low-keV virtual monoenergetic images (VMIs) can improve radiomics-based differentiation remains an important unanswered question (13,14).

Radiomics has emerged as an important quantitative imaging approach that enables the high-throughput extraction of mineable features from routine medical images, thereby extending radiological assessment beyond subjective visual interpretation. By quantifying lesion intensity, shape, texture, and spatial heterogeneity, radiomics can capture subvisual information potentially associated with tissue composition, vascular architecture, and microstructural complexity. Mariotti et al. emphasized that radiomics should be regarded as a structured multistep workflow involving image acquisition and preprocessing, segmentation, feature extraction and selection, and model training and validation, with careful attention to methodological rigor and reproducibility (15). Flaiban et al. further highlighted that radiomics is contributing to a shift in radiology from qualitative description toward quantitative analysis and supports the development of imaging biomarkers with potential diagnostic, prognostic, and predictive value (16). Within this framework, radiomics may offer incremental value for the non-invasive characterization of renal masses with overlapping conventional imaging appearances, particularly when subtle differences in intratumoral heterogeneity and microstructural organization are difficult to appreciate by visual assessment alone.

The integration of VMI with radiomics may provide a specific opportunity in renal mass evaluation because low-keV reconstructions may accentuate subtle vascular and intratumoral heterogeneity that is not readily appreciable on conventional images (17). At the same time, radiomic features are sensitive to reconstruction energy, image noise, and gray-level distribution, meaning that energy selection should be justified not only by visual contrast enhancement but also by its expected influence on quantitative feature behavior (18-20). In view of the contrast-enhancing advantages of low-keV imaging and the energy dependence of radiomic features, 45-keV VMIs were selected as a technically reasonable low-energy reconstruction for quantitative assessment of small hypervascular renal masses, although the radiomic stability of this specific energy level still warrants further validation. Thus, the application of low-keV VMI-based radiomics to renal fp-AML/ccRCC differentiation remains under-investigated and represents a meaningful knowledge gap.

Dual-layer spectral CT (DLCT) remains more costly and less accessible than conventional CT, particularly outside specialized referral centers (21,22). Nevertheless, if a spectral CT-based radiomics model improves the non-invasive characterization of equivocal small renal masses, it may still provide meaningful value in specialized oncologic imaging workflows rather than universal first-line deployment (23,24).

More broadly, machine learning has shown increasing value in renal oncology for tasks such as tumor subtype classification, mutation prediction, outcome stratification, and treatment-response assessment (25,26). Against this background, the novelty of the present study lies in integrating DLCT, dual-phase (CMP + NP) 45-keV VMI reconstruction, and radiomics-based machine learning to address the specific problem of distinguishing fp-AML from ccRCC among small renal masses. Building on prior evidence supporting the discriminative potential of CT radiomics, we aimed to evaluate whether a 45-keV VMI-based radiomics approach could facilitate the non-invasive characterization of these lesions by capturing subtle vascular and compositional differences. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0020/rc).

Methods

Figure 1 illustrates the radiomics workflow devised for the differential diagnosis of renal fp-AML and ccRCC, which includes data acquisition, VOI delineation, feature extraction, feature selection, and model training and validation.

Figure 1 Workflow of the radiomics study. Preoperative spectral CT data were retrospectively collected from patients with renal tumors. Tumor ROIs were manually segmented on 45-keV VMIs at the CMP and NP. Radiomics features were extracted from original, LoG, and wavelet-filtered images, including shape, intensity, and texture-based features. Feature selection involved normalization, statistical evaluation, and LASSO regression to construct the Rad-score. Model performance was evaluated using 10-fold CV, ROC curve analysis, AUC, DCA, and calibration curves. Model interpretation was conducted with cluster heatmaps, Pearson correlation analysis, and SHAP. AUC, area under the curve; CMP, corticomedullary phase; CT, computed tomography; CV, cross-validation; DCA, decision curve analysis; LASSO, least absolute shrinkage and selection operator; LoG, Laplacian of Gaussian; NP, nephrographic phase; Rad-score, radiomics score; ROC, receiver operating characteristic; ROIs, regions of interest; SHAP, Shapley additive explanations; VMI, virtual monoenergetic image.

Study design

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of Shenzhen Hospital, Southern Medical University (No. NYSZYYEC2025K100R001), and the requirement for informed consent was waived for this retrospective study. All data were retrieved from the Picture Archiving and Communication System (PACS) database, covering the period from January 2019 to June 2025. Eligible patients who met the inclusion and exclusion criteria during this period were consecutively enrolled. Because this was a retrospective single-center study, no formal prospective sample size calculation was performed. The final sample size was determined by the number of eligible consecutive patients available during the study period. Patients were randomly assigned to the training and validation cohorts at a 7:3 ratio. Given the limited sample size, a temporal split was not applied. The inclusion criteria were as follows: (I) all patients underwent triphasic renal CT acquisition, including the unenhanced phase (UP), corticomedullary phase (CMP), and nephrographic phase (NP), using the same CT scanning protocol; (II) pathological confirmation following partial or total surgical resection. The exclusion criteria were as follows: (I) visible fat on unenhanced CT images as determined by two radiologists; (II) tumors with a maximum diameter greater than 4 cm; (III) poor image quality with artifacts that hindered diagnostic interpretation. Accordingly, this study specifically focused on small renal masses (≤4 cm), as these lesions represent a clinically relevant setting in which preoperative differentiation between fp-AML and ccRCC remains particularly difficult. The final study population consisted of 108 eligible patients, including 50 patients with fp-AML and 58 patients with ccRCC. As both cohorts originated from the same institution and imaging platform, this approach provided internal validation only and could not substitute for independent external validation across centers or scanner protocols. Patient demographics, including age and sex, as well as tumor imaging features, including maximum diameter, unenhanced CT attenuation, tumor-renal cortex angle, cystic changes, pseudocapsule sign, and calcification, were collected from the electronic medical records. The training set was used for model construction and parameter tuning during cross-validation, whereas the validation set was used to evaluate model performance independently, without involvement in feature selection, feature standardization, or model construction.

CT image acquisition

All examinations were performed using a 64/128-slice dual-layer spectral CT (IQon spectral CT, Philips Medical Systems, Best, The Netherlands), covering both kidneys. The scanning protocol included data acquisition in three phases. Image acquisition was initiated using bolus tracking in the descending aorta at a threshold of 100 Hounsfield units (HU). CMP images were acquired 8 s after bolus triggering, followed by NP acquisition 65 s later. Patients received 100 mL of contrast agent (iodixanol, 350 mg/mL), which was injected into the antecubital vein at a rate of 3.0 mL/s using a high-pressure injector. The scanning parameters were as follows: tube voltage 120 kV, tube current 200 mA, collimation 64 mm × 0.625 mm, rotation time 0.75 s, pitch 1.016:1, and field of view 360 mm. Images were reviewed using a window width of 360 HU and a window level of 40 HU. Although unenhanced images were acquired for diagnostic assessment, radiomics analysis was performed on the 45-keV VMIs reconstructed from the CMP and NP. Image post-processing was performed using the Philips IntelliSpace Portal workstation.

CT image analysis

Two radiologists, with 5 and 3 years of abdominal diagnostic experience, respectively, independently analyzed the CT images without prior knowledge of the pathological results. The primary criteria for assessment included the presence of calcification, cystic changes, necrosis, and pseudocapsule, as well as measurements of the maximum tumor diameter and the angle between the tumor and renal cortex. Cystic changes were defined as well-defined, regularly shaped low-density areas without enhancement. In cases of disagreement between the two radiologists, consensus was reached through discussion.

Image preprocessing

Radiomics feature extraction, feature selection, and machine learning model construction were performed using the Darwin research platform (https://arxiv.org/abs/2009.00908), which is based on PyRadiomics and is consistent with the Image Biomarker Standardisation Initiative (IBSI). Before feature extraction, all 45-keV VMIs from the CMP and NP underwent standardized preprocessing. To reduce the impact of differences in acquisition parameters and spatial resolution on feature stability, all images were resampled to isotropic voxels of 1×1×1 mm³ using sitkBSpline interpolation. Gray-level discretization was performed using an equal-width binning method with a bin width of 25. Default platform intensity normalization was disabled, and voxel intensities were linearly scaled to a range of 0.0–4,096.0 after window adjustment to reduce inter-case differences in gray-level distribution.

CT image segmentation

Tumor segmentation was manually performed on the 45-keV CMP and NP images by two radiologists with 5 and 3 years of experience, respectively, who were blinded to the pathological results. A whole-lesion segmentation strategy was adopted, and the regions of interest (ROIs) included solid tumor components as well as necrotic and cystic areas to preserve intratumoral heterogeneity, better reflect the overall imaging phenotype, and reduce subjectivity associated with selective exclusion, particularly when the boundary between viable tumor tissue and non-enhancing components was indistinct. After manual delineation of the tumor boundary on each slice, three-dimensional reconstruction was performed to generate the volume of interest (VOI). During manual segmentation, care was taken to avoid inclusion of adjacent normal renal parenchyma to minimize partial-volume effects. Radiomic features were extracted from the original 45-keV VMIs and corresponding filtered images, including exponential, squared, square root, logarithmic, LoG-sigma-3.0-mm-3D, and wavelet-transformed images. A total of 1,781 features, including shape-based, first-order, and texture features, were acquired from each phase image. Segmentation reproducibility was evaluated using the intraclass correlation coefficient (ICC). Interobserver agreement was assessed based on independent segmentations by the two radiologists, whereas intraobserver agreement was assessed from repeated segmentation by one radiologist after a 2-week interval. Only features with an ICC >0.80 were retained for subsequent analysis. To avoid information leakage, all univariate screening and downstream feature selection procedures were performed exclusively within the training cohort.

Feature selection and radiomics signature construction

All radiomics features were normalized using the min-max scaling method, which transforms the values to a [0, 1] range, thereby reducing the influence of differences in feature magnitude. Dimensionality reduction was subsequently performed using the minimum redundancy maximum relevance (mRMR) method followed by least absolute shrinkage and selection operator (LASSO) regression. Five machine learning classifiers—random forest (RF), extreme gradient boosting (XGBoost), logistic regression (LR), support vector machine (SVM), and k-nearest neighbors (k-NN)—were developed for binary classification of fp-AML and ccRCC. These algorithms were chosen to represent commonly used modeling strategies with distinct learning mechanisms, including linear, margin-based, instance-based, bagging, and boosting approaches. To reduce the risk of overfitting in the setting of a relatively small sample size and high-dimensional feature space, dimensionality reduction was performed before model construction, model training was performed with 10-fold cross-validation in the training cohort, and the held-out validation cohort was kept fully independent from feature selection, scaling, and model fitting. Model performance was evaluated using ROC analysis, and calibration performance was visually assessed using calibration plots in the training and validation cohorts. Additionally, Shapley additive explanations (SHAP) values were incorporated during model development to enhance interpretability, and the output of the optimal model was visualized.

Statistical analysis

All statistical analyses were performed using R software (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were compared using the independent-samples t-test or Mann-Whitney U test, according to data distribution assessed by the Shapiro-Wilk test. Categorical variables were compared using the Chi-squared test or Fisher’s exact test, as appropriate. Model performance was assessed using area under the curve (AUC), accuracy (ACC), sensitivity (SEN), SPE, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Differences in AUCs between models were compared using the DeLong test. All tests were two-tailed, with statistical significance set at P<0.05.

Results

Patient characteristics

Our cohort comprised 108 patients (58 with ccRCC and 50 with fp-AML). Compared with ccRCC, fp-AML lesions exhibited significantly smaller diameters (P<0.001), higher unenhanced CT attenuation values (P<0.001), and a female predominance (P=0.01). Additionally, cystic degeneration and pseudocapsule formation were more frequent in ccRCC (both P<0.001). Detailed baseline data are summarized in Table 1.

SHAP-based interpretation of the optimal radiomics model

To enhance interpretability, SHAP values were computed for the optimal SVM model to reveal the relative influence of individual features on prediction results. The corresponding SHAP beeswarm plots illustrate the distribution of fp-AML and ccRCC cases based on key radiomic features.

CMP features

Six radiomic features showed significant differences in the CMP, as shown in Figure S1. Among these, squareroot_firstorder_Maximum exhibited the highest SHAP value, indicating that it may play an important role in distinguishing ccRCC from fp-AML. Higher values of this feature were more frequently observed in ccRCC. This correlation may reflect the hypervascular nature of ccRCC, which may demonstrate heterogeneous enhancement due to intratumoral angiogenesis, necrosis, and hemorrhage on CMP CT imaging. In particular, focal hyperenhancing regions in ccRCC during this phase may contribute to elevated first-order maximum values after square-root transformation.

NP features

Eight features differed significantly in the NP. The most discriminative feature was square_firstorder_Skewness, which was more strongly associated with ccRCC diagnosis. This metric is derived from the skewness of voxel intensities after square transformation, a preprocessing step that amplifies the weight of hyperdense pixels (e.g., enhancing tumor regions). In ccRCC, the resultant skewness may reflect asymmetric density distribution possibly associated with necrosis, fibrosis, or vascular anomalies—hallmarks of renal malignancy. Detailed information is provided in Figure S2.

Combined CMP + NP multiphase analysis

When both phases were combined, 10 features reached statistical significance, including four NP-derived features and six CMP-derived features (Figure 2). The top discriminator was exponential_glrlm_RunEntropy, a texture feature reflecting intratumoral heterogeneity. Malignant lesions exhibited higher RunEntropy values, possibly associated with their complex architecture (e.g., necrosis, irregular borders). Conversely, benign fp-AML cases showed lower values, concordant with their homogeneous enhancement patterns. However, this interpretation remains speculative and requires further radiologic-pathologic correlation.

Figure 2 SHAP summary plot for the dual-phase dataset highlighting the most influential radiomic features in the model. Features such as exponential_glrlm_RunEntropy_NP, square_glszm_GrayLevelVariance_NP, and lbp-3D-k_glrlm_RunLengthNonUniformityNormalized_CMP had the strongest contributions to model predictions. CMP, corticomedullary phase; NP, nephrographic phase; SHAP, Shapley additive explanations.

Evaluation of model performance

A total of 15 models were constructed using five machine learning algorithms, including XGBoost, SVM, LR, RF, and k-NN, under three feature-group settings. The classification performance metrics of these models are summarized in Figure 3, and the receiver operating characteristic (ROC) curves and decision curve analysis (DCA) are presented in Figure 4. Across different feature-group settings, the combined dual-phase model generally outperformed the single-group models. Detailed performance metrics are provided in Table S1. Overall, the SVM-based combined dual-phase model achieved the numerically best performance. In the training set, the AUC was 0.941 [95% confidence interval (CI): 0.885–0.996], F1 score =0.889, ACC =0.893, SEN =0.914, and SPE =0.875; in the validation set, the AUC was 0.915 (95% CI: 0.819–1.000), F1 score =0.848, ACC =0.848, SEN =0.867, and SPE =0.833. In addition, within the training cohort, all five classifiers were trained and tuned using 10-fold cross-validation. The combined dual-phase model showed consistently higher AUCs than the single-phase approach, with mean values of 0.87 for RF, 0.86 for SVM, 0.86 for XGBoost, 0.83 for k-NN, and 0.89 for LR. The aggregated AUC was 0.86±0.02, indicating the robust, algorithm-independent performance of dual-phase radiomic features (Table 2). Pairwise comparisons of ROC-AUCs using the DeLong test are summarized in Tables S2,S3. No statistically significant differences were found among the five classifiers within the same feature group in either the training or validation cohort (all P>0.05). Likewise, no significant differences were observed among the dual-phase, NP, and CMP models within the same classifier in either dataset (all P>0.05). However, in the validation set, the dual-phase model showed a trend toward higher AUC than the NP model for LR (ΔAUC =0.211, P=0.06) and SVM (ΔAUC =0.159, P=0.10), although these differences did not reach statistical significance. Calibration analysis was additionally performed to assess the agreement between predicted and observed probabilities. The calibration curve of the optimal dual-phase SVM model showed overall agreement with the ideal diagonal line in the validation cohort, although some deviation was observed in the low- to intermediate-probability range. Calibration plots for all candidate models are provided in Figure S3.

Figure 3 Performance comparison of five machine learning models across different datasets and phases. Radar plots illustrate the performance of LR, RF, SVM, XGBoost, and k-NN models in both training (top row) and validation (bottom row) sets. The datasets include dual-phase, CMP, and NP. Performance metrics include ACC, SEN, SPE, PRE, REC, F1-score, PPV, and NPV. ACC, accuracy; CMP, corticomedullary phase; k-NN, k-nearest neighbors; LR, logistic regression; NP, nephrographic phase; NPV, negative predictive value; PPV, positive predictive value; PRE, precision; REC, recall; SEN, sensitivity; SPE, specificity; SVM, support vector machine; XGBoost, extreme gradient boosting.

Figure 4 Comparison and evaluation of different models using include ACC, SEN, SPE, PRE, REC, F1-score, PPV, and NPV. ROC curves of each model in the training cohort (A-C) and testing cohort (D-F). DCA curves of each model in the training cohort (G-I) and testing cohort (J-L). The dual-phase model based on SVM demonstrated the highest discriminatory efficacy in the test cohort, achieving an AUC of 0.915. ACC, accuracy; AUC, area under the curve; CMP, corticomedullary phase; DCA, decision curve analysis; NP, nephrographic phase; NPV, negative predictive value; PPV, positive predictive value; PRE, precision; REC, recall; RF, random forest; ROC, receiver operating characteristic; SEN, sensitivity; SPE, specificity; SVM, support vector machine.

Discussion

Accurate differentiation between fp-AML and ccRCC is critical for clinical decision-making. In the present study, spectral CT-derived 45-keV VMIs were used to develop a 3D radiomics-based machine learning model, which demonstrated good diagnostic performance for discriminating fp-AML from ccRCC. The use of 45-keV VMIs may be advantageous because lower-keV reconstructions increase iodine attenuation and lesion conspicuity, potentially improving the detectability of subtle radiomic heterogeneity compared with higher-keV or conventional polychromatic images (27-30). Furthermore, comparative analysis of CMP, NP, and dual-phase strategies showed that the dual-phase approach achieved higher AUC and more balanced classification performance than the single-phase models.

The CMP primarily reflects arterial blood supply differences in tumors, typically demonstrating early abnormal hyperenhancement in ccRCC. Conversely, the NP highlights interstitial perfusion kinetics, exemplified by rapid contrast washout in necrotic areas or delayed enhancement in fibrous septa. The combined dual-phase modality has been validated as an effective radiomics strategy (31). Its core advantages lie in spatiotemporal feature complementarity, allowing the extraction of integrated perfusion-related biomarkers that are not available from single-phase imaging alone. Moreover, dual-phase imaging may improve anatomical depiction, leading to more accurate tumor boundary delineation and better visualization of morphological features, particularly pseudocapsules and lobulated contours. Furthermore, dual-phase feature fusion enables a more comprehensive assessment of intratumoral heterogeneity. Compared with single-phase approaches, dual-phase integration significantly expands tumor feature capture dimensionality, thereby improving machine learning model discriminative performance—a finding consistent with prior research. Previous studies have also demonstrated the advantages of combining CMP and NP. For instance, Kocak et al. (32) employed CMP to predict BAP1 mutation status in ccRCC, while Varghese et al. (33) and Cheng et al. (34) utilized both CMP and NP in contour-based segmentation, highlighting the superior classification outcomes achieved with this dual-phase strategy. Taken together, these findings suggest that the combination of CMP and NP may provide complementary information for differentiating fp-AML from ccRCC. However, given the retrospective single-center design and the lack of external validation, the clinical applicability of the present model should be interpreted cautiously. Although the proposed model may provide non-invasive quantitative support for lesion characterization, its broader clinical use will still depend on further validation and greater standardization of the imaging and analysis workflow.

Recent advancements in radiomics underscore the critical role of image preprocessing in extracting biologically relevant imaging biomarkers. As emphasized by Gillies et al. (35), applying advanced mathematical transformations (e.g., wavelets, Laplacian of Gaussian filters) accentuates intratumoral heterogeneity and reveals subtle textural variations linked to pathophysiologic processes such as hypoxia, necrosis, and altered cell density. These preprocessing steps enable the extraction of higher-order agnostic features beyond human visual perception, significantly enhancing the biological interpretability of radiomic signatures. Notably, exponential and other nonlinear transformations improve detection of poorly perfused or necrotic regions—features often associated with aggressive tumor phenotypes—reinforcing the necessity of standardized preprocessing protocols for predictive modeling in tumor characterization and treatment response assessment. However, prior studies differentiating fp-AML from ccRCC have predominantly relied on CT texture analysis (36), which faces inherent limitations in benign-malignant differentiation. Conventional approaches typically extract only second-order radiomic features, constraining their capacity to characterize complex spatial patterns including periodicity, directionality, morphological boundaries, and multiscale structural variations. This methodological limitation may be compounded by the narrow selection of machine learning models used in some previous studies, with most studies employing just one or two conventional algorithms (9,37). Nie et al. (38) found that the majority of the selected radiomics features (12/14) were higher-order wavelet and filter-based features, which go beyond the capabilities of traditional texture analysis. These advanced features provide a more nuanced representation of tumor heterogeneity, significantly enhancing the model’s ability to differentiate between hyperattenuating ccRCC and AML. Given that tumor heterogeneity is a critical factor in distinguishing benign from malignant lesions, incorporating higher-order radiomics features could offer additional diagnostic value by capturing more complex spatial information that second-order features alone may not reveal. In line with recent radiomics research, this study incorporated first-order, second-order, and higher-order features to comprehensively assess renal masses. The findings of this study further reveal that higher-order features predominantly contribute to the diagnostic significance. Specifically, five of the six features identified in the CMP were higher-order, while six out of eight features in the NP were of the same category. Among the 10 features selected from both phases, only one was not classified as a higher-order feature. SHAP-based interpretability analysis identified exponential_glrlm_RunEntropy from the NP as the most influential feature in the dual-phase contrast-enhanced CT radiomics model. This finding may be related to the relatively homogeneous enhancement of the renal parenchyma during the NP, which provides a more stable background for detecting subtle differences in intratumoral heterogeneity. In this context, entropy-related features may reflect variations in gray-level complexity associated with differences in vascular heterogeneity and microstructural organization between fp-AML and ccRCC. ccRCC typically shows rich but heterogeneous vascularity and more complex internal architecture, whereas fp-AML contains varying proportions of smooth muscle cells, abnormal vessels, and fat-poor components. These differences may contribute to distinct texture patterns on NP images.

Notably, Takahashi et al. evaluated entropy as a heterogeneity-related CT parameter and reported that it contributed to the contrast-enhanced CT model for differentiating fp-AML from RCC, supporting the potential value of entropy-related features in this setting (8). In addition, a recent systematic review and meta-analysis in oncologic CT texture analysis suggested that entropy and other heterogeneity-related texture features are repeatedly associated with clinically relevant tumor characteristics across studies, further supporting the broader relevance of entropy-based radiomic descriptors (39). However, in our study, exponential_glrlm_RunEntropy, rather than conventional entropy, emerged as the most influential feature. This may be because the exponential transformation increases SEN to subtle low-gray-level texture variations, thereby improving the detection of fine-scale heterogeneity related to enhancement uniformity and tissue microstructure. Nevertheless, this interpretation remains inferential and requires further radiologic-pathologic validation.

Erdim et al. (37) utilized eight machine learning methods to differentiate between benign and malignant renal tumors. The results demonstrated that the RF algorithm performed the best in distinguishing renal tumors, while other models exhibited generally lower predictive performance. Lee et al. (40) proposed a classification system based on texture features, combining three different feature selection techniques with four classifiers. The use of ReliefF for feature selection, combined with KNN and SVM classifiers, resulted in significantly improved diagnostic ACC (72.3% and 72.1%, respectively; P<0.05) for distinguishing fp-AML from ccRCC, outperforming other combinations of classifier and feature selection methods. SVM, renowned for its structural risk minimization principle and robust classification performance in small-sample, high-dimensional data scenarios, has progressively emerged as one of the cornerstone algorithms in radiomics research for renal tumors (40). In accordance with these findings, our study showed that, among the five classifiers, SVM achieved the best predictive performance. Compared with the study by Lee et al., our model achieved superior results in terms of ACC and AUC, possibly due to differences in sample balance and feature selection strategies. Notably, unlike the study by Erdim et al. (37), where most models showed relatively poor performance, all five classifiers in our study exhibited generally good performance. Furthermore, DeLong tests comparing ROC AUCs revealed no statistically significant differences among the five models. Although these findings support the robustness of the dual-phase radiomics strategy across different classifiers, caution is still warranted when considering its clinical applicability, given the retrospective single-center design and the lack of external validation.

This study has several limitations. First, this was a retrospective single-center study with random splitting rather than temporal splitting; thus, the current 7:3 training-validation strategy represents internal validation only and may have yielded optimistic performance estimates. External multicenter validation is still required. Second, the dual-phase protocol increases radiation exposure. Third, although reproducibility was evaluated, manual segmentation remains susceptible to observer-related variability, and inclusion of necrotic and cystic components may partly reflect secondary tumor changes. Finally, the model was developed using conventional handcrafted radiomics features and was restricted to lesions ≤4 cm; therefore, its applicability is limited to small renal masses.

Conclusions

In conclusion, our radiomic analysis demonstrated that a dual-phase spectral CT approach (dual-phase) using 45-keV virtual monoenergetic reconstructions showed better diagnostic performance for differentiating fp-AML from ccRCC than single-phase methods. This advantage was observed consistently across the five machine learning classifiers. These findings suggest that dual-phase spectral CT radiomics may represent a promising non-invasive approach for this challenging diagnostic task. Further multicenter validation is warranted to confirm its generalizability and potential clinical utility.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0020/rc

Data Sharing Statement: Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0020/dss

Peer Review File: Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0020/prf

Funding: This research was supported by the National Natural Science Foundation of China (No. 82471941), Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515220081), Guangdong Provincial Science and Technology Program (No. 2024A0505040005), Shenzhen Natural Science Foundation (Nos. JCYJ20250604183553066, JCYJ20230807142306014, and JCYJ20230807142308018), and the Shenzhen Postdoctoral Research Foundation (No. UN-KC-BHKY202204).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-1-0020/coif). All authors report the funding from the National Natural Science Foundation of China (No. 82471941), Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515220081), Shenzhen Natural Science Foundation (Nos. JCYJ20250604183553066, JCYJ20230807142306014, and JCYJ20230807142308018), and the Shenzhen Postdoctoral Research Foundation (No. UN-KC-BHKY202204). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of Shenzhen Hospital, Southern Medical University (No. NYSZYYEC2025K100R001), and the requirement for informed consent was waived for this retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Kratzer TB, Giaquinto AN, et al. Cancer statistics, 2025. CA Cancer J Clin 2025;75:10-45. [Crossref] [PubMed]
2. Larcher A, Campi R, Bex A, et al. Epidemiology of Renal Cancer: Incidence, Mortality, Survival, Genetic Predisposition, and Risk Factors. Eur Urol 2025;88:341-58. [Crossref] [PubMed]
3. Bhatt JR, Richard PO, Kim NS, et al. Natural History of Renal Angiomyolipoma (AML): Most Patients with Large AMLs >4cm Can Be Offered Active Surveillance as an Initial Management Strategy. Eur Urol 2016;70:85-90. [Crossref] [PubMed]
4. Dawidek MT, Villada JSA, Vazquez-Rivera K, et al. Higher Risk Renal Angiomyolipomas: Surveillance Remains a Safe Management Option. Urol Pract 2025;12:274-80. [Crossref] [PubMed]
5. Silvestri A, Gavi F, Sighinolfi MC, et al. Management of Small Renal Masses: Literature and Guidelines Review. Int Braz J Urol 2025;51:e20250203. [Crossref] [PubMed]
6. Frank I, Blute ML, Cheville JC, et al. Solid renal tumors: an analysis of pathological features related to tumor size. J Urol 2003;170:2217-20. [Crossref] [PubMed]
7. Finelli A, Ismaila N, Bro B, et al. Management of Small Renal Masses: American Society of Clinical Oncology Clinical Practice Guideline. J Clin Oncol 2017;35:668-80. [Crossref] [PubMed]
8. Takahashi N, Leng S, Kitajima K, et al. Small (< 4 cm) Renal Masses: Differentiation of Angiomyolipoma Without Visible Fat From Renal Cell Carcinoma Using Unenhanced and Contrast-Enhanced CT. AJR Am J Roentgenol 2015;205:1194-202. [Crossref] [PubMed]
9. Yan L, Liu Z, Wang G, et al. Angiomyolipoma with minimal fat: differentiation from clear cell renal cell carcinoma and papillary renal cell carcinoma by texture analysis on CT images. Acad Radiol 2015;22:1115-21. [Crossref] [PubMed]
10. Bellin MF, Valente C, Bekdache O, et al. Update on Renal Cell Carcinoma Diagnosis with Novel Imaging Approaches. Cancers (Basel) 2024;16:1926. [Crossref] [PubMed]
11. Wilson MP, Patel D, Murad MH, et al. Diagnostic Performance of MRI in the Detection of Renal Lipid-Poor Angiomyolipomas: A Systematic Review and Meta-Analysis. Radiology 2020;296:511-20. [Crossref] [PubMed]
12. Dehghani Firouzabadi F, Gopal N, Hasani A, et al. CT radiomics for differentiating fat poor angiomyolipoma from clear cell renal cell carcinoma: Systematic review and meta-analysis. PLoS One 2023;18:e0287299. [Crossref] [PubMed]
13. Coppola A, Tessitore L, Fontana F, et al. Dual-Energy Computed Tomography in Urological Diseases: A Narrative Review. J Clin Med 2024;13:4069. [Crossref] [PubMed]
14. Zhu Q, Sun J, Zhu W, et al. Spectral CT imaging versus conventional CT post-processing technique in differentiating malignant and benign renal tumors. Br J Radiol 2023;96:20230147. [Crossref] [PubMed]
15. Mariotti F, Agostini A, Borgheresi A, et al. Insights into radiomics: a comprehensive review for beginners. Clin Transl Oncol 2025;27:4091-102. [Crossref] [PubMed]
16. Flaiban E, Orhan K, Gonçalves BC, et al. Radiomics in Action: Multimodal Synergies for Imaging Biomarkers. Bioengineering (Basel) 2025;12:1139. [Crossref] [PubMed]
17. Patel BN, Farjat A, Schabel C, et al. Energy-Specific Optimization of Attenuation Thresholds for Low-Energy Virtual Monoenergetic Images in Renal Lesion Evaluation. AJR Am J Roentgenol 2018;210:W205-17. [Crossref] [PubMed]
18. Euler A, Laqua FC, Cester D, et al. Virtual Monoenergetic Images of Dual-Energy CT-Impact on Repeatability, Reproducibility, and Classification in Radiomics. Cancers (Basel) 2021;13:4710. [Crossref] [PubMed]
19. Wolf EV, Müller L, Schoepf UJ, et al. Photon-counting detector CT-based virtual monoenergetic reconstructions: repeatability and reproducibility of radiomics features of an organic phantom and human myocardium. Eur Radiol Exp 2023;7:59. [Crossref] [PubMed]
20. Zhong J, Pan Z, Chen Y, et al. Robustness of radiomics features of virtual unenhanced and virtual monoenergetic images in dual-energy CT among different imaging platforms and potential role of CT number variability. Insights Imaging 2023;14:79. [Crossref] [PubMed]
21. Rajiah P, Parakh A, Kay F, et al. Update on Multienergy CT: Physics, Principles, and Applications. Radiographics 2020;40:1284-308. [Crossref] [PubMed]
22. Toshav A. Economics of Dual-Energy CT: Workflow, Costs, and Benefits. Semin Ultrasound CT MR 2022;43:352-4. [Crossref] [PubMed]
23. Bellini D, Panvini N, Laghi A, et al. Systematic Review and Meta-Analysis Investigating the Diagnostic Yield of Dual-Energy CT for Renal Mass Assessment. AJR Am J Roentgenol 2019;212:1044-53. [Crossref] [PubMed]
24. Cao J, Lennartz S, Pisuchpen N, et al. Renal Lesion Characterization by Dual-Layer Dual-Energy CT: Comparison of Virtual and True Unenhanced Images. AJR Am J Roentgenol 2022;219:614-23. [Crossref] [PubMed]
25. Raman AG, Fisher D, Yap F, et al. Radiomics and Artificial Intelligence: Renal Cell Carcinoma. Urol Clin North Am 2024;51:35-45. [Crossref] [PubMed]
26. Durant AM, Medero RC, Briggs LG, et al. The Current Application and Future Potential of Artificial Intelligence in Renal Cancer. Urology 2024;193:157-63. [Crossref] [PubMed]
27. Nagayama Y, Nakaura T, Awai K, et al. Contrast medium dose optimization in the era of multi-energy CT. Jpn J Radiol 2025;43:1729-45. [Crossref] [PubMed]
28. Li G, Rednam N, Kundra V. Low KeV virtual monoenergetic images for detecting low dose iodine- or alternative Gd-based IV contrast agents. Heliyon 2024;10:e35210. [Crossref] [PubMed]
29. Chen FX, Jiang KK, Zhu JF, et al. Diagnostic accuracy of dual-layer spectral computed tomography virtual monoenergetic imaging with multiplanar reformation for T-staging of colorectal cancer. World J Gastroenterol 2025;31:110573. [Crossref] [PubMed]
30. Schabel C, Patel B, Harring S, et al. Renal Lesion Characterization with Spectral CT: Determining the Optimal Energy for Virtual Monoenergetic Reconstruction. Radiology 2018;287:874-83. [Crossref] [PubMed]
31. Song H, Wang X, Wu R, et al. The influence of manual segmentation strategies and different phases selection on machine learning-based computed tomography in renal tumors: a systematic review and meta-analysis. Radiol Med 2024;129:1025-37. [Crossref] [PubMed]
32. Kocak B, Durmaz ES, Kaya OK, et al. Machine learning-based unenhanced CT texture analysis for predicting BAP1 mutation status of clear cell renal cell carcinomas. Acta Radiol 2020;61:856-64. [Crossref] [PubMed]
33. Varghese B, Cen S, Zahoor H, et al. Feasibility of using CT radiomic signatures for predicting CD8-T cell infiltration and PD-L1 expression in renal cell carcinoma. Eur J Radiol Open 2022;9:100440. [Crossref] [PubMed]
34. Cheng D, Abudikeranmu Y, Tuerdi B. Differentiation of Clear Cell and Non-clear-cell Renal Cell Carcinoma through CT-based Radiomics Models and Nomogram. Curr Med Imaging 2023;19:1005-17. [Crossref] [PubMed]
35. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
36. Ma Y, Cao F, Xu X, et al. Can whole-tumor radiomics-based CT analysis better differentiate fat-poor angiomyolipoma from clear cell renal cell caricinoma: compared with conventional CT analysis? Abdom Radiol (NY) 2020;45:2500-7. [Crossref] [PubMed]
37. Erdim C, Yardimci AH, Bektas CT, et al. Prediction of Benign and Malignant Solid Renal Masses: Machine Learning-Based CT Texture Analysis. Acad Radiol 2020;27:1422-9. [Crossref] [PubMed]
38. Nie P, Yang G, Wang Z, et al. A CT-based radiomics nomogram for differentiation of renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma. Eur Radiol 2020;30:1274-84. [Crossref] [PubMed]
39. Zhang YR, Lu YH, Lin CM, et al. Pretreatment CT Texture Analysis for Predicting Survival Outcomes in Advanced Nonsmall Cell Lung Cancer Patients Receiving Immunotherapy: A Systematic Review and Meta-Analysis. Thorac Cancer 2025;16:e70144. [Crossref] [PubMed]
40. Lee HS, Hong H, Jung DC, et al. Differentiation of fat-poor angiomyolipoma from clear cell renal cell carcinoma in contrast-enhanced MDCT images using quantitative feature classification. Med Phys 2017;44:3604-14. [Crossref] [PubMed]