Skip to main content
NCBI home page
Future Oncology logoLink to Future Oncology
. 2025 Oct 21;21(28):3629–3638. doi: 10.1080/14796694.2025.2577631

The predictive value of dual-energy computed tomography radiomics in microvessel of clear cell renal cell carcinoma

Ruobing Li a,b, Xinyou Su c, Zhongyuan Li b,d, Ning Wang b, Haitao Sun b, Aimei Ouyang b,✉
PMCID: PMC12674250  PMID: 41120253

ABSTRACT

Aims

This retrospective study aimed to assess the potential of radiomic features extracted from dual-energy computed tomography (DECT) images, combined with machine learning algorithms, for the noninvasive prediction of microvessel density (MVD) in clear cell renal cell carcinoma (ccRCC).

Methods

We manually segmented regions of interest (ROIs) on corticomedullary phase (CMP) images to extract radiomic features. Tumor microvessel parameters were determined by immunohistochemical staining. Prediction models for MVD were developed using both multi-energy and monoenergetic sequence DECT images. Subsequently, a combined model was constructed based on the best-performing radiomics score and statistically significant clinical features, and was visualized as a nomogram. Furthermore, an external validation cohort was recruited from Center II to evaluate the performance of the nomogram.

Results

The support vector machine (SVM) classifier achieved the best performance for the multi-energy sequence MVD prediction model, with an AUC of 0.914 in the validation set. The MVD prediction model based on iodine-based material decomposition images (IMDI), constructed using the SVM classifier, achieved an AUC of 0.889 in the validation set. The nomogram showed good calibration, achieving an AUC of 0.757 in the external validation cohort.

Conclusions

DECT-based radiomic features show potential for noninvasive predicting microangiogenesis in patients with ccRCC.

KEYWORDS: Clear cell renal cell carcinoma, radiomics, microvessel density, imaging biomarker, dual-energy CT

Plain Language Summary

This study tested if a special kind of CT scan, called dual-energy CT, can help doctors know how many tiny blood vessels are in a common kidney cancer without needing surgery. We used computers to look at the scan pictures and patient information to make a system that guesses the density of these tiny blood vessels. One computer method we used worked well when we tried it on patients. We also made an easy tool that mixes the scan results and patient details to help doctors choose the best treatment for each person. This way, doctors can learn about the tumor without doing any painful tests, making care safer and better.

1. Introduction

In the United States alone, an estimated 81,800 new cases of kidney and renal pelvis cancers were diagnosed in 2023, resulting in approximately 16,701 deaths [1]. Renal cell carcinoma (RCC) accounts for approximately 85% of kidney tumors, with clear cell RCC (ccRCC) as the most common subtype, comprising roughly 70% of RCC cases [2]. Despite advances in the diagnosis and treatment of RCC, up to 17% of patients present with distant metastasis at the time of diagnosis, and the 5-year survival rate for these individuals remains below 20% [3]. This underscores the importance of early detection and accurate prognostic assessment.

Radiological heterogeneity within tumors is often influenced by alterations in blood flow [4,5]. Compared to normal tissue, malignant tumors have thinner, more tortuous, and irregular microvessel walls – features closely associated with tumor growth, invasion, and metastasis [6]. This has spurred increasing interest in noninvasive quantification of tumor-related angiogenesis. Quantification of microvessel is primarily assessed through indirect measurements, with microvessel density (MVD) being the most employed parameter. MVD is widely regarded as a reliable predictor of tumor progression and metastasis potential, and has been shown to correlate with the prognosis of RCC as well as various other malignancies [7–9]. Moreover, noninvasive assessment of MVD may facilitate more accurate identification of patients likely to benefit from anti-angiogenic therapies, offering valuable guidance for targeted treatment strategies, particularly those aimed at inhibiting angiogenesis [10].

Currently, imaging techniques such as computed tomography (CT) are widely employed in the evaluation of diseases affecting the urinary and reproductive systems. However, conventional CT imaging presents inherent limitations in tissue characterization and lesion differentiation [11]. The emergence of dual-energy computed tomography (DECT) has shifted imaging technology from monoenergetic to multi-energy acquisition strategies. By acquiring data at multiple different energy levels, DECT enables precise separation and microscopic characterization of lesions [12,13]. However, conventional image interpretation remains largely subjective and may fail to fully capture intratumoral heterogeneity. To address this, we incorporated radiomics techniques to facilitate quantitative image analysis [14]. As nearly all cancer patients undergo routine imaging examinations, these images represent a valuable resource for radiomics analysis. Radiomics-based machine learning algorithms can extract noninvasive imaging biomarkers that reflect the underlying physiological and pathological heterogeneity of the disease [15].

This study aimed to assess whether radiomic features reflecting tumor heterogeneity could serve as predictors of microvessel neogenesis in ccRCC, thereby providing valuable insights into the association between radiomic features and underlying tumor biology.

2. Methods

2.1. Study cohort

This is a retrospective study primarily analyzing patients’ medical imaging characteristics and clinical data, without involving personal privacy or direct health intervention. This study adhered to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines for reporting observational studies. The study strictly adheres to the ethical principles of the Declaration of Helsinki and has been approved by the Medical Ethics Committee of Central Hospital Affiliated to Shandong First Medical University, thus exempting the requirement for informed consent from patients.

We obtained an initial cohort of 123 RCC cases, confirmed by pathology, from our hospital’s imaging archive and exchange system. The inclusion criteria were as follows: I ccRCC confirmed by histopathology after radical or partial nephrectomy; II complete DECT-enhanced scans performed within one month before surgery; and III single lesions only, to avoid the influence of multiple lesions on the delineation of the regions of interest (ROIs). Exclusion criteria were shown in Figure 1. A total of 87 ccRCC-confirmed patients were randomly divided into a training set of 69 cases and a validation set of 18 cases, following an 8:2 ratio. Additionally, 17 patients from Centre II were recruited as an independent external validation cohort. Demographic information, laboratory results, and clinical characteristics were extracted from the electronic medical record system, including gender, age, nuclear grade, T-stage, tumor maximum diameter, lipid levels, blood glucose, neutrophil ratio, eosinophil ratio, urea nitrogen, creatinine, albumin, albumin/globulin ratio, lactate dehydrogenase, and serum potassium. Cases with incomplete imaging or clinical data were excluded from analysis.

Figure 1.

Figure 1.

Open in a new tab

The flowchart of the study.

2.2. DECT imaging protocol

CT image acquisition followed a standardized protocol using a dual-energy contrast-enhanced DECT scanner (Somatom Force; Siemens Healthcare, Forchheim, Germany). All patients received an intravenous injection of a nonionic contrast agent (300 mg iodine/ml) at a dose of 1.2 mL/kg body weight, administered at a flow rate of 3.5 mL/s. Scanning was performed in automatic exposure and dual-energy modes, with the corticomedullary phase (CMP) and nephrogenic phase (NP) delayed by 30 and 80 seconds, respectively. Tube voltages were set at 100 kVp and sn150 kVp, with effective tube currents of 80–90 mA and 130–180 mA, respectively. Post-processes of the acquired images using the ADMIRE reconstruction algorithm of the workstation (syngo.via) to generate iodine-based material decomposition images (IMDI), monoenergetic 70 keV, 100 keV, and 150 keV, mixed energy images (MEI) and virtual non-contrast (VNC) images [16].

2.3. Tumor segmentation and features extraction

All images were uploaded to Radcloud (Huiying Medical Technology Co., Ltd, https://www.huiyihuiying.com) via the Uploaer software for image management and data preprocessing. Two radiologists, each with over 8 years of experience in abdominal imaging diagnosis, semi-automatically delineated the lesion regions on axial CMP images to obtain ROI. All contours were subsequently reviewed by a senior radiologist. The clinical data involved in this study were anonymized, and the annotators were aware that the lesions were ccRCC but were blinded to all other clinical information. The ROIs included tumor tissue, necrosis, fibrosis, and hemorrhagic areas, while normal renal parenchyma was excluded.

Pre-processing such as image resampling and gray scale discretization were performed on the images prior to extraction of radiomics features according to the Image Biomarker Standardization Initiative (IBSI) [17]. The images and ROIs are resampled using linear interpolation to ensure that the voxel size of all images is standardized to 1 mm × 1 mm × 1 mm, achieving isotropic voxels. Additionally, Z-score normalization is applied to standardize the image grayscale, effectively minimizing variations in grayscale values caused by differences in scanning equipment or parameters. Feature selection was conducted solely on the training set. The feature types were classified into three categories: I first-order statistical features, which describe the distribution of voxel intensities; II size- and shape-based features, which reflect the size and shape of the region; and III texture features, which quantify the heterogeneity within the region. Texture features included the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM).

2.4. Microvessel parameter acquisition

The surgical specimens of ccRCC were sliced into 4 μm thick sections, fixed in formalin, and embedded in paraffin. The sections were then immunohistochemical staining for CD34 (ZSGB-Bio rabbit #ZA-0550). A pathologist with 5 years of experience conducted assessments of the sections using a polarized light microscope (Olympus BX53, Tokyo, Japan) at magnifications of 40×, 100×, and 200 × . Initial assessments were performed at lower magnifications (40×, 100×) to identify tumor areas and regions with higher vascular density, and to assess microvessel morphology. Subsequently, five microscopic photos were taken at 200× magnification for each sample, resulting in a total of 435 images. Quantitative analysis of CD34+ stained areas was performed using Image-Pro Plus 6.0 software (Media Cybernetics, Silver Spring, MD, USA). Tumor microvessel assessment was based on the Weidner method [18], which uses CD34-labeled endothelial cells as the standard for microvessel counting. Modifications were made as needed, and the final MVD was calculated as the ratio of the total number of immunopositivity clusters to the total area.

2.5. Radiomics models construction

We used Radcloud (http://radcloud.cn/) for radiomics statistical analysis. The models optimize feature subset construction through the following steps: I variance threshold: set a threshold of 0.8 to exclude features with variance below this value, removing low-variance features to reduce noise, II SelectKBest: features with p-values less than 0.05 were selected, and III the least absolute shrinkage and selection operator (LASSO): First, LASSO regression analysis is performed using the cv.glmnet function, with L1 regularization for feature selection, helping to eliminate redundant features and prevent model overcomplication. Next, 5-fold cross-validation is used to assess the model’s generalizability, and the optimal regularization parameter lambda is automatically selected to further control model complexity. Additionally, by limiting the maximum number of iterations to 1000 and standardizing the data before training (ensuring each feature has a mean of 0 and a standard deviation of 1), the model’s stability is enhanced.

Six machine learning classifiers were employed in this study, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Logistic Regression (LR), and Decision Tree (DT). A multi-energy sequence MVD prediction model was constructed based on six energy images and the aforementioned classifiers. Classifier performance was evaluated by Precision (P), Recall (R), and F1 Score (P × R ×2/(P + R)). The best-perform classifier was subsequently employed to develop MVD prediction models based on six monoenergetic sequence images. The predictive effectiveness of the radiomics models was assessed by the area under the receiver operating characteristic curve (AUC). Decision curve analysis (DCA) was obtained performed the “rmda” package to evaluate the clinical utility of the models.

2.6. Nomogram construction

Single-factor and multifactor logistic regression analyses were performed on clinical and pathological features to identify factors significantly affecting event probability (P < 0.05). A nomogram was generated to predict MVD classification based on the optimal Radiomics score (Rad-score) and these feature parameters. WHO/ISUP grade was evaluated as a potential prognostic factor. The nomogram’s performance was assessed using AUC, DCA, and calibration curves were generated with the “rms” package to evaluate the model’s net benefit across various probability thresholds.

2.7. Statistical analysis

The flowchart of this study is shown in Figure 1. Categorical variables were compared using the chi-square test, while continuous variables were assessed using the Mann-Whitney U test. A P-value < 0.05 was considered statistically significant. All data analyses were conducted using R Software (version 4.2.1, https://www.r-project.org/) and Statistical software version 26 (IBM, Armonk, NY).

3. Results

3.1. Patient characteristics

A total of 87 eligible ccRCC patients (60 males and 27 females) with a mean age of 57 (±10.94) years were finally enrolled in Centre I. The median MVD was 387. There were no significant differences in clinical features and microvessel parameters between the training cohort and the validation cohort (Table 1). Center II included 17 eligible ccRCC patients (11 male, 6 female) with a mean age of 56 years ±10.31 years as an external validation cohort for the nomogram.

Table 1.

Clinical and pathological characteristics of study participants with ccRCC in the training and validation sets.

Variable Training set (n = 69) Validation set (n = 18) P-value
Gender     0.813
 Male (%) 48 (69.6) 12 (66.7)  
 Female (%) 21 (30.4) 6 (33.3)  
Age (years) 57.38 ± 11.395 56.17 ± 9.556 0.818
Fuhrman grade     0.737
 I, II, I – II (%) 51 (73.9) 14 (77.8)  
 III, IV, III – IV (%) 18 (26.1) 4 (22.2)  
Microvessel parameter     0.956
 High-MVD (%) 35 (50.7) 9 (50)  
 Low-MVD (%) 34 (49.3) 9 (50)  
Open in a new tab

3.2. Features extraction

Initially, 10,128 radiomic features were acquired and categorized into three groups. Group 1 (First-order statistics) included 756 descriptors that quantified the distribution of voxel intensities in CT images using basic standard metrics. Group 2 (Shape and size-based features) comprised 84 3D features that reflecting the shape and size characteristics of the regions. Group 3 (Texture features) included of 3150 texture features calculated using Gray Level Run Length Matrix (GLRLM) and Gray Level Co-occurrence Matrix (GLCM), which quantified heterogeneity changes in the regions.

3.3. Classifier and radiomics models performance

The multi-energy sequence MVD prediction model built using the SVM classifier had the best performance (Figure 2(a,b)). In the training set, it achieved an AUC of 0.959 (95% CI: 0.889–1.000), with a sensitivity of 0.890 and specificity of 0.910. In the validation set, it achieved an AUC value of 0.914 (95% CI: 0.734–1.000), with a sensitivity of 0.890 and specificity of 0.780. DCA indicated that the multi-energy sequence MVD prediction model provided better clinical utility within a reasonable threshold probability range (Figure 2(c, d)).

Figure 2.

Figure 2.

Open in a new tab

Receiver operating characteristic curves of (A) training set and (B) internal validation set of multi-energy sequence MVD prediction models (from left-to-right: KNN, SVM, XGBoost, RF, LR and DT classifiers). Decision curve analysis of (C) training set and (D) internal validation set of multi-energy sequence MVD prediction models.

The performance metrics of various classifiers in the multi-energy sequence MVD prediction models are listed in Table 2. After eliminating classifiers that exhibited overfitting, the SVM classifier showed the best performance. In the validation set, it achieved an accuracy of 0.800, a recall of 0.890, and an F1-score of 0.840.

Table 2.

AUC values, 95% confidence intervals, sensitivity, specificity of the validation set of MVD prediction models constructed using six classifiers, along with Precision, Recall, F1-score for each classifier.

Classifier AUC (95% CI) Sensitivity Specificity Precision Recall F1-score
KNN 0.722
(0.542–0.902)		 0.890 0.780 0.800 0.890 0.840
SVM 0.914
(0.734–1.000)		 0.890 0.780 0.800 0.890 0.840
XGBoost 0.858
(0.652–1.000)		 0.780 0.780 0.780 0.780 0.780
RF 0.827
(0.647–1.000)		 0.890 0.780 0.800 0.890 0.840
LR 0.765
(0.585–0.945)		 0.890 0.780 0.800 0.890 0.840
DT 0.722
(0.512–0.932)		 0.780 0.670 0.780 0.780 0.740
Open in a new tab

Abbreviations: KNN: k-nearest neighbor, SVM: support vector machine, XGBoost: extreme gradient boost, RF: random forest, LR: logistic regression, and DT: decision tree.

The monoenergetic sequence MVD prediction model, built using 9 features extracted from IMDI, demonstrated the best performance (Figure 3(a, b)). In the training set, it achieved an AUC of 0.947 (95% CI: 0.869–1.000), with a sensitivity of 0.890 and a specificity of 0.850. In the validation set, the AUC was 0.889 (95% CI: 0.697–1.000), with a sensitivity of 0.670 and a specificity of 0.890. DCA indicated that the monoenergetic sequence prediction model offered clinical utility within a reasonable range of threshold probabilities (Figure 3(c,d)).

Figure 3.

Figure 3.

Open in a new tab

Receiver operating characteristic curves of (A) training set and (B) internal validation set of monoenergetic sequence MVD prediction models (from left-to-right: IMDI, 70 keV, 100kVp, 150kVp, MEI and VNC). Decision curve analysis of (C) training set and (D) internal validation set of monoenergetic sequence MVD prediction models.

3.4. Nomogram model performance

Based on the MVD classification of ccRCC patients, 44 (50.6%) patients were classified as high-MVD. These patients were more likely to have elevated blood lipids, urea nitrogen, and creatinine levels (P < 0.05) compared to low-MVD patients (Table 3). After multifactor logistic regression analysis, relevant features and the optimal Rad-score (derived from the multi-energy sequence model) were selected to construct a combined nomogram (Figure 4(a)). The ROC curve (Figure 4(b-d)), calibration curve (Figure 4(e-g)), and decision curve analysis (DCA) results (Figure 4(g–i)) of the nomogram are all displayed. In the internal validation cohort, the nomogram achieved an AUC of 0.800, demonstrating strong performance in the training cohort. In the external validation cohort, the AUC was 0.757, confirming the nomogram’s robustness and predictive accuracy on independent data. Calibration curves show good agreement between predictions and observations.

Table 3.

Patient characteristics of clinical factors according to high- and low-grade ccRCC subgroups.

Characteristic High-grade(n = 22) Low-grade(n = 65) Univariate analysis P-value multivariate analysis P-value
Age 56.32 ± 11.72 57.40 ± 10.83 0.6886  
Male sex* 18 42 0.1397  
Stage T3, 4* 3 18 0.1928  
Maximum diameter 51.78 ± 23.86 46.23 ± 20.51 0.2949  
Shape vesicular* 14 29 0.1272  
Glucose (3.9–6.1 mmol/L) 5.52 ± 1.28 5.57 ± 1.33 0.8649  
Blood lipids (0.3–1.7 mmol/L) 1.94 ± 1.54 1.38 ± 0.72 0.0415 0.0861
Neutrophil ratio (40–75%) 67.66 ± 11.77 68.67 ± 11.76 0.7224  
Eosinophil ratio (0.4–8%) 2.00 ± 1.69 1.72 ± 1.89 0.5413  
Urea nitrogen (2.30–7.80 mmol/l) 6.48 ± 1.13 5.20 ± 1.40 0.0012 0.0042
Creatinine (62-115umol/l) 99.05 ± 42.61 79.86 ± 21.90 0.0139  
Albumin (40-55 g/L) 43.33 ± 4.75 43.32 ± 4.94 0.9909 0.1258
Albumin/Globulin (1.2–2.4) 2.71 ± 4.54 2.02 ± 2.33 0.3816  
Lactate dehydrogenase (120-230 U/L) 203.82 ± 39.95 194.85 ± 41.30 0.3749  
Blood potassium (3.5–5.3 mmol/L) 4.15 ± 0.28 4.07 ± 4.29 0.4235  
Open in a new tab

Data were means ± SDs (SD Standard deviation).

*Data presented were the number of patients.

Figure 4.

Figure 4.

Open in a new tab

The nomogram (A), combining clinical factors and Rad-score. Receiver operating characteristic curves of the nomogram in the (B) training set, (C) internal validation set, and (D) external validation set. Calibration curves of the nomogram in the (E) training set, (F) internal validation set, and (G) external validation set. The 45° straight line indicates the ideal prediction; the closer the two curves are, the higher the accuracy. Decision curve analysis of the nomogram in the (H) training set, (I) internal validation set, and (J) external validation set.

4. Discussion

The results showed a correlation between radiomic features and microvessel growth in ccRCC, with the radiomic features from multi-energy sequences and the IMDI model performing well in predicting microvessel. Additionally, the nomogram model for predicting MVD demonstrated good agreement with actual observations.

DECT is becoming increasingly common in clinical practice, providing better characterization of renal masses and improved contrast for renal parenchyma, vessels, and lesions. Studies by Bellini et al. [19] and Salameh et al. [20] have demonstrated the high accuracy of DECT in evaluating renal masses. Additionally, Zhang et al. [6] explored the predictive value of DECT quantitative parameters in RCC microangiogenesis and found that parameters such as mean iodine concentration (IC), mean water concentration (WC), mean normalized IC and mean normalized WC were positively correlated with microvessel parameters, but only during the CMP. Our pre-experimental results also show that models constructed using NP images features have slightly lower AUC values than those based on CMP. According to previous studies and our findings, using only CMP-based radiomics analysis does not affect prediction accuracy.

Previous studies have shown that multi-energy sequence models perform better than monoenergetic sequence models. Al Ajmi et al. [21] extracted texture features of tumors from virtual monoenergetic imaging (VMI) ranging from 40 keV to 140 keV and compared them with VMIs at 65 keV, demonstrating that multi-energy dataset texture analysis is superior to single-energy dataset texture analysis. Separately, Xu et al. [22] integrated large-scale radiological and radiomics features to identify reliable factors for predicting microvessel invasion with the constructed model demonstrating strong predictive performance (AUC of 0.889 for the test sets). Our study also found that the multi-energy sequence model outperformed the monoenergetic sequence models, likely because multi-energy sequence images contain more detailed information about lesion features.

The results of this study showed that the radiomic features extracted in the multi-energy sequence model were primarily from IMDI, while the performance of the IMDI-based monoenergetic sequence model was the best. Therefore, we concluded that IMDI is more suitable for assessing tumor microvessel. The IMDI generated by DECT minimizes artifacts, enhances lesion contrast, and allows for a more direct estimation of tumor vascularity and angiogenesis [23]. Relevant literature supports this finding. Patel et al. [24] found that using uniform normalized iodine thresholds reduces variability by quantifying aortic iodine, enabling better differentiation between vascular and nonvascular lesions. Regarding VMI, we found that lower VMI energy levels result in better predictive model performance. Lennartz et al. [25] discussed the clinical application of DECT images in the abdomen, noting that low-energy VMI enhances vascular assessment, as lower energy levels approach the iodine absorption maximum, providing an angiography-like view in a single-phase examination. However, when VMI energy levels exceed 80 keV, tissue contrast decreases. For MEI, the presence of redundant data may result in poor image features. Similarly, VNC image features performed poorly, likely because blood vessels are identified primarily based on iodine content, and VNC images remove this information, reflecting only tumor characteristics [26,27].

In recent years, radiomics techniques to study the tumor microenvironment have primarily focused on histological features and molecular markers. The main challenges include the invisibility of early tumor diagnosis and the lack of positive examination indicators. The concept of using quantitative image data combined with specific characteristics to represent different regions of lesion physiology has been widely validated, showing great potential for disease diagnosis, differentiation, and prediction in clinical practice [28,29]. The development of novel machine learning algorithms requires continuous optimization, coordination, and standardization of the entire process based on research results, while also allowing for modifications and evolution of methods relevant to the current state of radiomics [30,31]. In this study, microvessel density was selected as a predictive endpoint primarily because microvessels are a crucial component of the tumor microenvironment. Many tumors, including ccRCC, require neovascularization to support growth, invasion, and eventual metastasis. ccRCC is one of the most vascularized tumor types [32,33]. Existing studies have confirmed the correlation between malignant tumor progression and microvessel parameter levels [34,35]. We chose to use CD34, which has been shown to be particularly well-expressed in the presence of increased tumor neovascularization and is most strongly correlated with microvessel expression [36,37]. Additionally, subjective differences in microvessel counts and areas may have led to varying results. To minimize this, we used professional image software for automated counting.

There are certain limitations with this study. First, as a retrospective analysis, it is subject to selection bias, and future prospective studies are needed to confirm the robustness of the models. Secondly, the small sample size – particularly that of the external validation set – may limit the model’s generalizability and reduce the statistical power to detect performance variability. Future research should validate the model in larger, independent, and multi-center cohorts to ensure its reliability and applicability across different clinical settings. Additionally, longer follow-up periods are necessary to assess long-term outcomes more comprehensively.

5. Conclusion

Our study demonstrates that DECT-based radiomic features are associated with microvessel growth in ccRCC, which could serve as a diagnostic marker for microvessels in ccRCC and help identify physiological processes related to vascularity. This provides a reference for optimizing individualized therapeutic regimens and improving patient prognosis.

Funding Statement

This work was supported by the Jinan Municipal Health Commission under [Grant number 202019036] and the Department of Science and Technology of Shandong Province under [Grant number YDZX2021012].

Article highlights

Objective

  • To evaluate the utility of radiomic features from dual-energy CT (DECT) for noninvasive prediction of tumor microvessel characteristics in clear cell renal cell carcinoma (ccRCC).

Reference Standard

  • Tumor microvessel density (MVD) was quantified using histopathological analysis.

Modeling Approach

  • Predictive models for MVD were developed using machine learning algorithms, particularly support vector machines (SVM).

Key Imaging Data

  • Radiomic features were extracted from both multi-energy and monoenergetic sequence DECT images.

Model Performance

  • The SVM model based onmulti-energy DECT images achieved high predictive accuracy; theIMDI-based model also demonstrated strong performance.

Combined Model

  • A nomogram was constructed by integrating the most predictive radiomic features with significant clinical variables.

Clinical Utility

  • The nomogram facilitated individualized prediction of tumor microangiogenesis and showed good performance in external validation.

Conclusion

  • DECT-based radiomics offers a promising noninvasive tool for assessing tumor vascular characteristics in ccRCC.

Author contributions

All the authors contributed to the content development and the writing of this study and approved the final draft.

R. Li: Data curation, Formal analysis, Validation, Visualization, Writing – original draft, Writing – review & editing. X. Su: Data curation, Formal analysis, Writing – review & editing, Supervision. Z. Li: Formal analysis, Validation, Writing – review & editing, Supervision. N. Wang: Investigation, Methodology, Writing – review & editing. H. Sun: Methodology, Conceptualization, Writing – review & editing, Funding Provided. A. Ouyang: Methodology, Supervision, Project administration, Writing – review & editing, Funding Provided.

Ethical declaration

This retrospective study protocol was reviewed and approved by the Medical Ethics Committee of Central Hospital Affiliated to Shandong First Medical University, approval number 2023–012-01. The study has been granted an exemption from requiring written informed consent.

Disclosure statement

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.

Reviewer disclosure

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Data availability statement

The data that support the findings of this study are available from the corresponding author, (A. Ouyang), upon reasonable request.

References

Papers of special note have been highlighted as either of interest (•) or of considerable interest (••) to readers.

  • 1.Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA A Cancer J Clinicians. 2023;73(1):17–48. doi: 10.3322/caac.21763 [DOI] [PubMed] [Google Scholar]
  • 2.Motzer RJ, Jonasch E, Agarwal N, et al. Kidney cancer, version 3.2022, NCCN clinical practice guidelines in oncology. J Natl Compr Cancer Network. 2022;20(1):71–90. doi: 10.6004/jnccn.2022.0001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Capitanio U, Bensalah K, Bex A, et al. Epidemiology of renal cell carcinoma. Eur Urol. 2019;75(1):74–84. doi: 10.1016/j.eururo.2018.08.036 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Hinshaw DC, Shevde LA.. The tumor microenvironment innately modulates cancer progression. Cancer Res. 2019;79(18):4557–4566. doi: 10.1158/0008-5472.CAN-18-3962 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Gatenby RA, Grove O, Gillies RJ. Quantitative imaging in cancer evolution and ecology. Radiology. 2013;269(1):8–14. doi: 10.1148/radiol.13122697 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Zhang B, Wu Q, Qiu X, et al. Effect of spectral CT on tumor microvascular angiogenesis in renal cell carcinoma. BMC Cancer. 2021;21(1):874. doi: 10.1186/s12885-021-08586-x [DOI] [PMC free article] [PubMed] [Google Scholar]; • This study demonstrated that spectral CT-derived quantitative parameters, particularly iodine- and water-based metrics, correlate significantly with microvascular density and area in RCC, supporting the use of dual-energy CT as a noninvasive biomarker of tumor angiogenesis.
  • 7.Mertz KD, Demichelis F, Kim R, et al. Automated immunofluorescence analysis defines microvessel area as a prognostic parameter in clear cell renal cell cancer. Hum Pathol. 2007;38(10):1454–1462. doi: 10.1016/j.humpath.2007.05.017 [DOI] [PubMed] [Google Scholar]
  • 8.Marcon J, Graser A, Horst D, et al. Papillary vs clear cell renal cell carcinoma. Differentiation and grading by iodine concentration using DECT—correlation with microvascular density. Eur Radiol. 2020;30(1):1–10. doi: 10.1007/s00330-019-06298-2 [DOI] [PubMed] [Google Scholar]
  • 9.Xiao J, Rahbar H, Hippe DS, et al. Dynamic contrast-enhanced breast MRI features correlate with invasive breast cancer angiogenesis. NPJ Breast Cancer. 2021;7(1):42. doi: 10.1038/s41523-021-00247-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Ferician AM, Ferician OC, Nesiu A, et al. The mutually mediated chloride intracellular channel protein 1 (CLIC1) relationship between malignant cells and tumor blood vessel endothelium exhibits a significant impact on tumor angiogenesis, progression, and metastasis in clear cell renal cell carcinoma (ccRCC). Cancers (basel). 2022;14(23):5981. doi: 10.3390/cancers14235981 [DOI] [PMC free article] [PubMed] [Google Scholar]; •• This study identifies CLIC1 as a novel biomarker expressed in both tumor cells and endothelial cells in ccRCC, revealing a bidirectional interaction that actively promotes angiogenesis and metastasis. It establishes a mechanistic link between microvascular characteristics and tumor aggressiveness.
  • 11.Posada Calderon LP, Eismann L, Reese SW, et al. Advances in imaging-based biomarkers in renal cell carcinoma: a critical analysis of the current literature. Cancers (Basel). 2023;15(2):354. doi: 10.3390/cancers15020354 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Mileto A, Marin D. Dual-energy computed tomography in genitourinary imaging. Radiol Clin N Am. 2017;55(2):373–391. doi: 10.1016/j.rcl.2016.10.006 [DOI] [PubMed] [Google Scholar]
  • 13.Zhang C, Wang N, Su X, et al. Force dual‑energy CT in pathological grading of clear cell renal cell carcinoma. Oncol Lett. 2019. doi: 10.3892/ol.2019.11022 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–577. doi: 10.1148/radiol.2015151169 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Liu X, Elbanan MG, Luna A, et al. Radiomics in abdominopelvic solid-organ oncologic imaging: current status. Am J Roentgenol. 2022;219(6):985–995. doi: 10.2214/AJR.22.27695 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Zhou W, Schornak R, Michalak G, et al. Determination of optimal image type and lowest detectable concentration for iodine detection on a photon counting detector-based multi-energy CT system. Proc SPIE Int Soc Opt Eng. 2018;10573:105734U. doi: 10.1117/12.2294949 [DOI] [PMC free article] [PubMed] [Google Scholar]; •• This study systematically evaluates image types and detection thresholds for iodine in photon-counting CT systems, providing essential technical insights into optimal image selection.
  • 17.Fornacon-Wood I, Mistry H, Ackermann CJ, et al. Reliability and prognostic value of radiomic features are highly dependent on choice of feature extraction platform. Eur Radiol. 2020;30(11):6241–6250. doi: 10.1007/s00330-020-06957-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Weidner N. Current pathologic methods for measuring intratumoral microvessel density within breast carcinoma and other solid tumors. Breast Cancer Res TR. 1995;36(2):169–180. doi: 10.1007/BF00666038 [DOI] [PubMed] [Google Scholar]; • This seminal work was among the first to propose and validate microvessel density (MVD) as a surrogate marker for tumor aggressiveness and metastatic potential, establishing a methodological standard for histological quantification.
  • 19.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. Am J Roentgenol. 2019;212(5):1044–1053. doi: 10.2214/AJR.18.20625 [DOI] [PubMed] [Google Scholar]
  • 20.Salameh J-P, McInnes MDF, McGrath TA, et al. Diagnostic accuracy of dual-energy CT for evaluation of renal masses: systematic review and meta-analysis. Am J Roentgenol. 2019;212(4):W100–W105. doi: 10.2214/AJR.18.20527 [DOI] [PubMed] [Google Scholar]
  • 21.Al Ajmi E, Forghani B, Reinhold C, et al. Spectral multi-energy CT texture analysis with machine learning for tissue classification: an investigation using classification of benign parotid tumours as a testing paradigm. Eur Radiol. 2018;28(6):2604–2611. doi: 10.1007/s00330-017-5214-0 [DOI] [PubMed] [Google Scholar]
  • 22.Xu X, Zhang H-L, Liu Q-P, et al. Radiomic analysis of contrast-enhanced CT predicts microvascular invasion and outcome in hepatocellular carcinoma. J Hepatol. 2019;70(6):1133–1144. doi: 10.1016/j.jhep.2019.02.023 [DOI] [PubMed] [Google Scholar]
  • 23.Mileto A, Sofue K, Marin D. Imaging the renal lesion with dual-energy multidetector CT and multi-energy applications in clinical practice: what can it truly do for you? Eur Radiol. 2016;26(10):3677–3690. doi: 10.1007/s00330-015-4180-7 [DOI] [PubMed] [Google Scholar]
  • 24.Patel BN, Vernuccio F, Meyer M, et al. Dual-energy CT material density iodine quantification for distinguishing vascular from nonvascular renal lesions: normalization reduces intermanufacturer threshold variability. Am J Roentgenol. 2019;212(2):366–376. doi: 10.2214/AJR.18.20115 [DOI] [PubMed] [Google Scholar]; •• This study addresses the variability in iodine quantification across different DECT platforms and demonstrates that normalization to the aorta significantly enhances reproducibility and diagnostic performance.
  • 25.Lennartz S, Hokamp NG, Kambadakone A. Dual-energy CT of the abdomen: radiology in training. Radiology. 2022;305(1):19–27. doi: 10.1148/radiol.212914 [DOI] [PubMed] [Google Scholar]
  • 26.Graser A, Johnson TRC, Hecht EM, et al. Dual-energy CT in patients suspected of having renal masses: can virtual nonenhanced images replace true nonenhanced images? Radiology. 2009;252(2):433–440. doi: 10.1148/radiol.2522080557 [DOI] [PubMed] [Google Scholar]
  • 27.Xiao JM, Hippe DS, Zecevic M, et al. Virtual unenhanced dual-energy CT images obtained with a multimaterial decomposition algorithm: diagnostic value for renal mass and urinary stone evaluation. Radiology. 2021;298(3):611–619. doi: 10.1148/radiol.2021192448 [DOI] [PubMed] [Google Scholar]
  • 28.Tataru OS, Marchioni M, Crocetto F, et al. Molecular imaging diagnosis of renal cancer using 99mTc-Sestamibi SPECT/CT and Girentuximab PET-CT-current evidence and future development of novel techniques. Diagnostics. 2023;13(4):593. doi: 10.3390/diagnostics13040593 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Monjaras-Avila CU, Lorenzo-Leal AC, Luque-Badillo AC, et al. The tumor immune microenvironment in clear cell renal cell carcinoma. IJMS. 2023;24(9):7946. doi: 10.3390/ijms24097946 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Xia T, Zhou Z, Meng X, et al. Predicting microvascular invasion in hepatocellular carcinoma using CT-based radiomics model. Radiology. 2023;307(4):222729. doi: 10.1148/radiol.222729 [DOI] [PubMed] [Google Scholar]
  • 31.Kumar V, Gu Y, Basu S, et al. Radiomics: the process and the challenges. Magn Reson Imag. 2012;30(9):1234–1248. doi: 10.1016/j.mri.2012.06.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Mier JW. The tumor microenvironment in renal cell cancer. Curr Opin Oncol. 2019;31(3):194–199. doi: 10.1097/CCO.0000000000000512 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Safali M, Karslioğlu Y, Arpaci F, et al. A distinct microvascular pattern accompanied by aggressive clinical course in breast carcinomas: a fact or a coincidence? Pathol - Res Pract. 2010;206(2):93–97. doi: 10.1016/j.prp.2009.10.004 [DOI] [PubMed] [Google Scholar]
  • 34.Jardim JF, Galvis MM, Fabelo IR, et al. Intratumoral lymphatic vascular density is an independent factor for disease-free and overall survival in advanced stage oral squamous cell carcinoma. Oral Surg Oral Med Oral Pathol Oral Radiol. 2021;132(5):580–588. doi: 10.1016/j.oooo.2021.07.014 [DOI] [PubMed] [Google Scholar]
  • 35.Kang Z-P, Wang L-X, Liu J, et al. Expression of CEACAM1 and CD105 in renal cell carcinoma and its correlation with microvessel density. Crit Rev Eukaryot Gene Expr. 2021;31(1):1–9. doi: 10.1615/CritRevEukaryotGeneExpr.2020037168 [DOI] [PubMed] [Google Scholar]
  • 36.Liang H, Hu C, Lu J, et al. Correlation of radiomic features on dynamic contrast-enhanced magnetic resonance with microvessel density in hepatocellular carcinoma based on different models. J Int Med Res. 2021;49(3):030006052199758. doi: 10.1177/0300060521997586 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Tamma R, Ingravallo G, Annese T, et al. Tumor microenvironment and microvascular density in human glioblastoma. Cells. 2022;12(1):11. doi: 10.3390/cells12010011 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author, (A. Ouyang), upon reasonable request.

Articles from Future Oncology are provided here courtesy of Taylor & Francis

RESOURCES