CT radiomics for predicting postoperative disease specific survival in clear cell renal cell carcinoma: a multi-center study

Introduction

Renal cell carcinoma (RCC), which originates from the epithelial cells of renal tubules, accounts for 80–90% of kidney malignancies, with clear cell RCC (ccRCC) being the predominant histopathological subtype (1). Diagnostic and therapeutic advancements have notably improved outcomes for patients with ccRCC. However, roughly 30% of patients experience recurrence and metastasis after radical nephrectomy (2). This underscores the need for precise prognostic assessments to tailor effective treatment plans. The primary clinical prognostic tools for ccRCC are the tumor, node, metastasis (TNM) staging system established by the American Joint Committee on Cancer and the nuclear grading system established by the World Health Organization/International Society of Urological Pathology (WHO/ISUP) (3,4). However, variability in survival rates among patients at the same stage reveals the limitations of these current models in fully capturing the tumor’s biological and prognostic indicators (5). Consequently, there is an urgent need for research focused on discovering effective prognostic markers to enhance the accuracy of current predictive models.

Radiomics, an emerging field, utilizes quantitative feature extraction from medical imaging to analyze tumor heterogeneity and biological characteristics (6). This approach effectively identifies subtle structural differences within tumors, revealing their heterogeneity and complexity—pivotal to advancing precision medicine (7). Within RCC research, radiomics has shown promise for diverse applications, including differentiating between benign and malignant tumors, accurate TNM staging, and WHO/ISUP grading (7-10). Additionally, radiomics has been utilized to identify gene expression patterns specific to RCC (11). Nonetheless, employing radiomics for ccRCC prognostic assessment faces challenges. Initial studies have not consistently achieved the desired accuracy levels in radiomics-based prognostic assessments, partly due to limited data sources and small sample sizes, which, in turn, limit the models’ generalizability. Furthermore, studies exploring the correlation between macroscopic radiomics features and microscopic pathomics features—quantitative features extracted from whole slide images (WSIs)—in ccRCC are scarce. The prognostic implications of integrating data across multiple scales remain largely unexplored.

This study utilized multi-center datasets to construct a robust, comprehensive ccRCC computed tomography (CT)-radiomics prognostic model, aiming to improve and supplement existing clinical assessment tools. We initially explored the correlation between radiomics and pathomics features, while assessing the prognostic utility of a multiscale approach that integrates clinical, radiological, and pathological data to enhance prognostic accuracy. We present this article in accordance with the TRIPOD + AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2284/rc).

Methods

Patient information

The research received approval from the Ethics Committee of Guizhou Provincial People’s Hospital (GZPH) (ethics No. [2021]110). Due to its retrospective design, the need for patient informed consent was waived in line with relevant ethical guidelines. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This multi-center study encompassed four datasets, each composed of ccRCC cases confirmed through postoperative pathology. These included cases from the Affiliated Hospital of Guizhou Medical University (GZMU) (July 2012–June 2020), which served as the training dataset, cases from GZPH (August 2013–June 2018), which formed the first external validation dataset, cases from the Affiliated Hospital of Zunyi Medical University (ZYMU) (July 2013–December 2017), which constituted the second external validation dataset, and cases from the C4KC-KITS collection hosted on The Cancer Imaging Archive (www.cancerimagingarchive.net), which acted as the third external validation dataset (12). All patients across these datasets were subject to the same inclusion and exclusion criteria. Specifically, inclusion criteria required surgical treatment, pathological confirmation of ccRCC, and CT-enhanced scans performed within 30 days prior to surgery. Exclusion criteria encompassed the absence of standard renal CT cortical phase enhanced images, any prior kidney interventions before the CT scan, poor CT image quality, incomplete clinical or pathological data, the presence of other concurrent tumors, and patients who died from causes other than ccRCC. The primary outcome measure of this study, termed disease-specific survival (DSS), was defined as the time from diagnosis to death due to ccRCC. The flow chart is shown in Figure 1.

Figure 1 The flow chart of the study. AIC, Akaike information criterion; ccRCC, clear cell renal cell carcinoma; C-index, concordance index; CT, computed tomography; IBS, integrated Brier score; ROI, region of interest; WSI, whole slide image.

Clinical, follow-up, radiological, and pathological data collection

In this study, a systematic approach was taken to gather clinical and pathological data from patients, encompassing details such as age, gender, tumor size, nuclear grade according to the WHO/ISUP classification, and TNM staging. The primary outcome measure of this study was DSS. Follow-up was conducted according to the National Comprehensive Cancer Network (NCCN) Kidney Cancer Guidelines. For patients managed before 2015, the 2012 guidelines recommended postoperative follow-up for ccRCC every 6 months for the first 2 years for American Joint Committee on Cancer TNM stages I–III, with annual follow-up thereafter, whereas stage IV follow-up was individualized (13). The 2015 guidelines revised this protocol, recommending follow-up for stage I patients every 6 months for 2 years, followed by annually; for stage II–III patients every 3–6 months for 3 years, followed by annually; and for stage IV patients every 6–16 weeks, with adjustments based on clinical judgment (14). The final follow-up cut-offs were 30 April 2023, for GZMU, and 30 August 2021, for both GZPH and ZYMU. CT cortical phase enhanced images from each center were obtained through the Picture Archiving and Communication System, ensuring standardized image retrieval across all participating institutions. For the GZMU dataset, pathologist Pinhao Li digitized representative tumor sections on hematoxylin and eosin-stained slides by utilizing the Leica Aperio GT450 scanner (Leica Biosystems, Nussloch, Germany) and preserved the images in .SVS format. For ccRCC patients in the C4KC-KITS dataset, The Cancer Genome Atlas (TCGA, https://www.cancer.gov/) was utilized to gather clinical, pathological, and DSS data. The Cancer Imaging Archive data center provided the CT images. In this study, all included patients possessed complete clinical records, pathological reports, and CT data, thereby ensuring the reliability and consistency of the analysis.

Tumor segmentation and radiomics feature extraction

Tumor segmentation in CT cortical phase enhanced images was conducted using ITK-SNAP software (version 3.8.0; https://www.itksnap.org/pmwiki/pmwiki.php). Experienced radiologists, Chong Tian and Yu Rong, manually outlined the tumor lesions in three dimensions to determine volumes of interest (VOIs). The tumor boundaries were delineated on each CT image slice, creating a three-dimensional (3D) volumetric representation. To assess the consistency of radiomics feature extraction, the intraclass correlation coefficient (ICC) was calculated for intra- and inter-observer variability. We randomly selected 10 patients from each dataset, totaling 40 patients, for this process. Radiologists Chong Tian and Yu Rong. independently delineated the VOI to evaluate the inter-observer ICC. Additionally, Chong Tian repeated the delineations 2 weeks later to assess intra-observer ICC.

The procedure began with resampling each CT image to standardize voxel size to 1×1×1 mm³. This standardization aids in ensuring uniform feature extraction. Using the Pyradiomics library (version 3.0.1), 107 radiomics features were initially extracted, covering shape, intensity, and texture categories, following the guidelines of the Image Biomarker Standardization Initiative (15). Additionally, by applying filtering techniques and wavelet transformations to the original images, an additional 1,395 features were derived. In total, 1,502 radiomics features were extracted for further analysis. Further details on these parameters can be found at (https://pyradiomics.readthedocs.io/en/latest/customization.html).

Feature selection and radiomics model development

The reliability of the extracted features was assessed through intra- and inter-observer ICC calculations, retaining only those with an ICC above 0.75 for further analysis. Subsequently, a detailed selection was performed using the least absolute shrinkage and selection operator (LASSO) Cox regression model. This approach was chosen for its efficacy in identifying the subset of radiomics features that offered the greatest predictive accuracy. Subsequently, a radiomics model was established using a linear combination of the selected features and their respective coefficients, which facilitated the computation of a radiomics score (Rad-score). Using the training cohort, an optimal Rad-score threshold was identified, allowing for the stratification of patients into low- and high-risk groups. This stratification threshold was then applied to the validation cohorts, confirming the model’s applicability and predictive validity across different patient populations.

Construction of clinical model and fusion model

TNM staging, the most widely used prognostic indicator in clinical practice, was selected as the clinical model. Additionally, a fusion model was developed by combining the TNM staging and the Rad-score, presented as a nomogram to enhance clarity and ease of use in predicting outcomes.

Model evaluation

The prognostic models, developed using the training dataset, were validated on external datasets. During validation, the model was neither updated nor recalibrated; its performance was directly evaluated on the validation datasets using the training set model. Evaluation metrics included concordance index (C-index) for assessment, with the Akaike information criterion (AIC) examining model complexity and fit. Furthermore, predictive accuracy was gauged using the integrated Brier score (IBS). Additionally, time-dependent receiver operating characteristic (ROC) analyses were performed at 1-, 3-, and 5-year intervals to evaluate the model’s ability to predict DSS. These analyses were accompanied by assessments of the model’s discrimination.

Pathological WSI processing and feature extraction

WSI is a technique that involves scanning traditional glass slides to produce digital images for analysis. For the GZMU dataset, an experienced pathologist, Pinhao Li, meticulously annotated the regions of interest (ROIs) on WSI using ImageScope software (Leica Biosystems). The ROI areas were precisely identified at level 0 resolution from each WSI and systematically segmented into 256×256 pixel blocks. These blocks underwent a color space conversion from red green blue (RGB) to Hue, Saturation, and Value (HSV). Subsequently, a foreground binary mask was created using median filtering and saturation threshold adjustment. Blocks predominantly consisting of background were excluded based on a 50% area threshold, ensuring a focus on blocks pertinent to pathological analysis. The number of image blocks extracted from each WSI varied considerably, ranging from a few hundred to several thousand, depending on the WSI’s dimensions.

For pathomics feature extraction, a convolutional neural network (CNN) was employed, specifically a pre-trained ResNet50 model augmented with ImageNet weights. Following the third residual block, an adaptive average pooling layer was introduced to transform each 256×256 block into a 1,024-dimensional feature vector. These vectors, derived from all blocks within a WSI, were then concatenated to form an n×1,024-dimensional matrix. Subsequently, the mean of each column was computed to yield a consolidated feature vector, serving as a representative of the entire WSI. As a result, each WSI was condensed into a single 1×1,024-dimensional vector, encapsulating the case’s pathomics features and providing a comprehensive source of pathological data for subsequent analyses.

Analysis and modeling of radiomics-pathomics feature correlations

An investigation was conducted to explore the correlations between pathomics features and patients’ DSS using a univariate Cox regression model. This approach enabled the quantification of the predictive power of individual pathomics features, as evidenced by their C-index scores. Following this analysis, the top 20 pathomics features with the highest predictive strength were carefully selected.

Furthermore, Spearman correlation was used to examine the associations between selected pathomics features and three key radiomics features, with the results visualized in a heatmap for clarity.

Additionally, the study focused on assessing the prognostic value of pathomics features, particularly those showing weak correlations with radiomics features. This assessment employed a multivariate Cox regression model, incorporating clinical and radiomics data as covariates. Finally, a multi-scale radiomics and pathomics (radiopathomics) model integrating clinical, radiomics, and pathomics data was developed for ccRCC prognostic assessment. Its effectiveness was evaluated using C-index, AIC, IBS, and time-dependent ROC analysis at 1-, 3-, and 5-year intervals, verifying its predictive accuracy for ccRCC prognostication.

Statistical analysis

The software R (version 3.6.1, R Foundation for Statistical Computing, Vienna, Austria) and Python (version 3.11.2, Python Software Foundation, Wilmington, DE, USA) were used for statistical analysis. For continuous variables, the t-test or Mann-Whitney U test was applied, depending on the data distribution, whereas categorical variables were assessed using either the Chi-squared test or Fisher’s exact test. Kaplan-Meier and log-rank tests were used for DSS estimation and survival curve comparison. Radiomics feature selection and prognostic model construction were performed using R’s “glmnet” and “survival” packages. Model accuracy was assessed with Harrell’s C-index, IBS, and AIC using “pec” and “SurvMetrics” packages. The “timeROC” and “rms” packages were used for time-dependent ROC analysis at 1-, 3-, and 5-year intervals, as well as for developing the calibration curve and nomogram. Spearman correlation was used to analyze variable relationships. The coxph function from the “survival” package was used for univariate and multivariate Cox regression analyses. Python’s scipy.stats, seaborn, and matplotlib libraries were employed for detailed variable analysis and heatmap visualization. P values <0.05 were considered significant.

Results

Clinical and pathological characteristics

A total of 530 ccRCC patients were included in the study after applying strict inclusion and exclusion criteria. The GZMU dataset functioned as the training dataset, comprising 219 patients and recording a mortality rate of 12.79% (28 deaths). For external validation, three additional datasets were used: GZPH with 174 patients and a 12.64% mortality rate (22 deaths), ZYMU with 86 patients and a 17.44% mortality rate (15 deaths), and the C4KC-KITS dataset with 51 patients and a 21.57% mortality rate (11 deaths). Across all datasets, survival analyses did not reveal any significant differences in DSS (P=0.30). Across the four centers, statistically significant differences (P<0.01) were observed in variables such as T stage, M stage, TNM stage, and WHO/ISUP grading, whereas other variables showed no significant differences (P>0.05). These differences are likely due to the distinct data distribution of the C4KC-KITS dataset compared to the other centers. A comprehensive summary of these characteristics is provided in Table 1.

Radiomics model construction

Initially, 1,502 radiomics features were extracted from CT cortical phase images of ccRCC patients. After excluding 49 features with intra- and inter-observer ICCs below 0.75, 1,453 features remained. Using the training dataset, a LASSO Cox regression algorithm identified three features significantly associated with DSS in ccRCC: one shape feature (original_shape_Maximum3DDiameter) and two texture features (square_gldm_DependenceNonUniformityNormalized, wavelet-LLH_glrlm_RunVariance). Using these pivotal features and their corresponding coefficients, a radiomics model was constructed to calculate each patient’s Rad-score.

The Rad-score formula is as follows:

Radiomicsscore=0.138885365042198×original_shape_Maximum3DDiameter+0.0169743546524579×square_gldm_DependenceNonUniformityNormalized+0.0192624672800098×wavelet-LLH_glrlm_RunVariance[1]

An optimal cutoff value of 1.045 was determined for this model. In the training dataset, the model exhibited a hazard ratio (HR) of 3.778 [95% confidence interval (CI): 2.378–6.003, P<0.001]. Survival analysis, based on the radiomics model and depicted through Kaplan-Meier curves, indicated a significant discriminative capability between low- and high-risk groups (log-rank test, P<0.001), as illustrated in Figure 2.

Figure 2 Comparative Kaplan-Meier survival curves for different models. (A) Training dataset, (B) validation dataset 1, (C) validation dataset 2, and (D) validation dataset 3.

Clinical and fusion model development

The clinical model, TNM staging, demonstrated an HR of 2.858 (95% CI: 2.048–3.989, P<0.001) in the training dataset. A fusion model was developed by combining TNM staging with the Rad-score. This fusion model showed an HR of 4.305 (95% CI: 2.754–6.730, P<0.001). Multivariate Cox regression analysis indicated that both the TNM staging and the Rad-score served as independent prognostic indicators for DSS in ccRCC patients (both P<0.001). The Kaplan-Meier survival curves illustrating the performance of both the clinical model and the fusion model can be found in Figure 2. It is worth noting that all log-rank test P values for the fusion model fell below 0.001, emphasizing its exceptional ability to differentiate between low- and high-risk patients.

Model evaluation

The fusion model achieved higher C-index values than the clinical and radiomics models across all datasets. Specifically, the fusion model’s C-index values were 0.804, 0.855, 0.830, and 0.814 across the datasets, outperforming both the clinical model (0.759, 0.815, 0.724, 0.802) and the radiomics model (0.794, 0.813, 0.803, 0.751), as detailed in Table 2. During the validation process, the model underwent no updates or recalibrations. The performance of the model, as developed on the training set, was directly evaluated on the validation datasets.

In external validation dataset 3, the fusion model had slightly higher IBS and AIC values than the clinical model, indicating marginally higher prediction error and a slightly worse balance between complexity and fit. However, in the training dataset and external validation dataset 2, the fusion model outperformed both the clinical and radiomics models with lower IBS and AIC values, and in external validation dataset 1 it achieved a lower AIC than the clinical model but a slightly higher IBS, indicating better predictive accuracy and generalization across most datasets. Despite this minor shortfall in external validation dataset 3, overall, the fusion model demonstrated superior performance, as shown in Table 2.

Across all datasets, the fusion model exhibited strong accuracy in predicting 1-, 3-, and 5-year DSS, as illustrated in Figure 3. Additionally, the fusion model is visually represented in the nomogram shown in Figure 4A. Calibration curves further validate the model’s predictive power, demonstrating strong agreement between the fusion model’s forecasts for 1-, 3-, and 5-year DSS rates and the actual observed survival rates, as shown in Figure 4. Comprehensive evaluation through C-index, IBS, AIC, and calibration curves confirmed its consistent and outstanding predictive capabilities across multiple datasets.

Figure 3 Time-dependent ROC curves comparing the predictive performance of fusion and radiopathomics models for 1-, 3-, and 5-year DSS rates. (A) Fusion model’s performance on the training dataset, (B) Radiopathomics model’s performance on the same dataset, and (C-E) fusion model’s performance on three validation datasets: validation dataset 1, validation dataset 2, and validation dataset 3, respectively. AUC, area under the curve; CI, confidence interval; DSS, disease-specific survival; ROC, receiver operating characteristic.

Figure 4 Fusion model nomogram and calibration curves. (A) Nomogram. (B-E) Calibration curves for 1-, 3-, and 5-year DSS in the (B) training dataset, (C) validation dataset 1, (D) validation dataset 2, and (E) validation dataset 3. DSS, disease-specific survival; TNM, tumor, node, metastasis.

Correlation analysis

Utilizing a pre-trained ResNet50 model with adaptive average pooling and a feature matrix compression strategy, 1,024 pathomics features were extracted from each WSI. Among these features, the top 20 demonstrated optimal predictive accuracy using a univariate Cox regression, with C-index values ranging from 0.707 to 0.763, highlighting the prognostic potential of pathomics features.

Spearman correlation analysis identified a notable pattern of correlation between radiomics and pathomics features, with pathomics features 168 and 266 being exceptions. Three key radiomics features exhibited moderate correlations with 18 prognostically relevant pathomics features, with correlation coefficients ranging from −0.417 to 0.386, visually represented in a heatmap (Figure 5).

Figure 5 Spearman correlation heatmap of radiomics and pathomics features, with correlation coefficients ranging from −0.42 to 0.39. Pathomics feature 341 exhibits the strongest negative correlation (correlation coefficients from −0.12 to −0.42), whereas pathomics feature 387 demonstrates the most significant positive correlation (correlation coefficients from 0.24 to 0.39). However, pathomics features 168 and 266 do not show any significant correlation with any radiomics feature. *, statistically significant correlations (P<0.05). PF, pathomics features; RF, radiomics features; RF 1, original_shape_Maximum3DDiameter; RF 2, square_gldm_DependenceNonUniformityNormalized; RF 3, wavelet-LLH_glrlm_RunVariance.

The radiopathomics model, integrating TNM staging, radiomics, and pathomics features 168 and 266, was shown to be a potent prognostic indicator (HR =11.019, 95% CI: 5.867–20.995, P<0.001). This model has significantly improved prognostic accuracy, achieving a C-index of 0.884, which surpasses the predictive capabilities of the clinical model (C-index: 0.759), radiomics model (C-index: 0.794), and even the fusion model (C-index: 0.804) on the training dataset (Table 2). Based on multivariate Cox regression analysis, the TNM staging (P=0.001), Rad-score (P=0.007), and pathomics features 168 and 266 (P<0.001) were all identified as independent prognostic factors. The radiopathomics model excelled in various performance metrics, including the AIC, IBS, and time-dependent ROC assessments for 1-, 3-, and 5-year outcomes (Table 2 and Figure 3). This underscores the significant prognostic value of the pathomics features in enhancing the predictive accuracy of the other models.

Discussion

This multi-center study developed a fusion model combining TNM staging and Rad-score to predict DSS in patients with ccRCC. The fusion model demonstrated superior accuracy compared to single models. Furthermore, the study uncovered a correlation between radiomics and prognostic pathomics features. A multiscale radiopathomics model, integrating uncorrelated prognostic pathomics features, Rad-score, and TNM staging, further improved the predictive accuracy of the fusion model for ccRCC DSS.

A prognostic radiomics model for ccRCC was developed using CT cortical phase images, focusing on three key features: the tumor’s maximum 3D diameter and two texture features. These texture features expose the tumor’s heterogeneity within CT scans, a property that has been associated with tumor aggressiveness and prognosis in previous studies (16). However, the morphological feature, particularly the tumor’s maximum 3D diameter, demonstrated a substantially higher coefficient in the Rad-score formula compared to the texture features, indicating its dominant role in prognostic prediction. This finding aligns well with the established importance of tumor size in TNM staging, where it serves as a crucial factor linked to tumor aggressiveness and patient outcomes (17-19).

TNM staging is a crucial prognostic factor for ccRCC, according to NCCN guidelines. Radiomics, extracting high-dimensional tumor features from CT imaging, offers valuable insights (20). However, their predictive accuracy varied across datasets in this study. Consequently, a fusion model combining TNM staging and Rad-score was developed. Multivariate Cox regression confirmed both as independent prognostic factors. The fusion model demonstrated superior accuracy and robustness in discrimination, fit, and complexity across multiple datasets, thereby enhancing prognostic assessment for ccRCC.

The results of this study exhibit several significant advantages compared to previous research. Previous studies have typically predicted overall survival rather than DSS, which focuses on deaths due to ccRCC, offering a more precise prognosis. Additionally, unlike He et al.’s use of a single prognostic evaluation metric, this study employed comprehensive prognostic evaluation metrics, demonstrating the superior accuracy and performance of the fusion model (21). In small ccRCC cohorts, fusion models based on radiomics have attained remarkable C-index scores for predicting postoperative prognosis (22,23). Nonetheless, these promising outcomes are tempered by limited generalizability stemming from small, single-center datasets. To address these constraints, this study employed a large-scale dataset from four diverse centers representing various ethnicities for extensive validation. Echoing both Nie et al.’s and Gao et al.’s multi-center ccRCC prognosis research, the fusion model excels in predicting both the DSS (via C-index) and specific 1-, 3-, and 5-year survival rates, while maintaining a good model fit (24-26).

This study further explored the relationship between prognostic radiomics and prognostic pathomics features in ccRCC. It uncovered a substantial correlation between these two types of features, suggesting a possible underlying connection via common biological mechanisms (27,28). This reinforces previous studies demonstrating radiomics’ ability to predict histopathological markers, such as nuclear grade and sarcomatoid change, in ccRCC (10,29,30). Nevertheless, the study identified the pathomics features 168 and 266, that lack correlation with radiomics, highlighting radiomics’ inability to capture certain microscopic prognostic details. This underscores the necessity of adopting a multiscale approach to obtain comprehensive prognostic information in ccRCC (27,31). Multivariate Cox analysis confirmed the independent prognostic significance of radiomics, pathomics, and clinical factors in ccRCC survival. Consistent with studies by Ning et al. and Schulz et al., the multi-scale radiopathomics model developed in this study outperformed single-scale models (32,33). This reflects the capacity of multi-scale features to capture tumor heterogeneity, thereby improving predictive accuracy and advancing the understanding of tumor behavior (27,34). These findings underscore the interconnectedness of radiomics and pathomics features, emphasizing the distinct prognostic value of both modalities.

This study developed fusion and radiopathomics models that significantly enhance the accuracy of clinical prognosis assessment, thereby improving treatment decision-making for patients with ccRCC. For low-risk patients, these models identify individuals with reduced recurrence and mortality risks, minimizing unnecessary medical interventions and avoiding treatment-related toxicity, particularly in those with impaired renal function (35,36). In high-risk patients, the models facilitate earlier and more precise therapeutic interventions, such as targeted therapy and immunotherapy, leading to improved treatment efficacy and survival rates (35,37). Additionally, the models optimize the allocation of medical resources by directing care to patients with the greatest clinical need, thereby enhancing overall healthcare efficiency.

Several limitations of this study should be noted, which also represent promising directions for future improvements. First, although this study used traditional radiomics and pathomics pipelines, advances in artificial intelligence, particularly deep learning, offer strong potential in automated, multi-level feature extraction (38). Future approaches using CNNs could act as “equation estimators” to capture complex, non-linear relationships potentially beyond hand-crafted features, potentially enhancing predictive performance (39). Second, feature reproducibility remains a challenge. The absence of dedicated test-retest or phantom datasets limits direct assessment of feature stability, and high-dimensional feature spaces increase the risk of false discoveries (40-42). Although multi-center data improve validation robustness, the retrospective design may still introduce selection bias. Robust feature selection, prospective external validation, harmonization, and standardized test-retest datasets will be essential for reliable clinical translation. Third, although correlations between prognostic radiomics and pathomics features were examined, the biological underpinnings—such as links to tumor heterogeneity, immune microenvironment, and signaling pathways—remain unclear. Clarifying these mechanisms will improve interpretability and clinical utility. Finally, although manual VOI delineation improves accuracy, automated segmentation techniques could boost efficiency and reduce operator variability, representing a promising direction for future advancements.

Conclusions

This multi-center study underscores the crucial role of radiomics in prognostic evaluations of ccRCC and its synergistic effect with clinical models. Additionally, it uncovers the correlation and complementarity between radiomics and pathomics features, enhancing prognostic assessments. The results emphasize the significance of radiomics in ccRCC analysis, paving the way for novel approaches to multiscale data integration in tumor prognosis.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD + AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2284/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2284/dss

Funding: The study was supported by the National Natural Science Foundation of China (Nos. 82160349, 82460344, and 82060314), the Guizhou Senior Innovative Talent Project (No. QKHPTRC-GCC[2022]041-1), Guizhou Province High-Level Innovative Talents ‘Thousand Levels’ Personnel Training Project (No. GZSYQCC[2023]004), Natural Scientific Project of Guizhou Province: Qiankehe Foundation - ZK[2022] General 263, Guizhou Provincial Health Commission Science and Technology Fund Project (Nos. gzwkj 2003-213, and gzwjkj 2019-1-202).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2284/coif). The authors have no 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. The research received approval from the Ethics Committee of Guizhou Provincial People’s Hospital (GZPH) (ethics No. [2021]110). Due to its retrospective design, the need for patient informed consent was waived in line with relevant ethical guidelines. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

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. Gray RE, Harris GT. Renal Cell Carcinoma: Diagnosis and Management. Am Fam Physician 2019;99:179-84.
2. Capitanio U, Montorsi F. Renal cancer. Lancet 2016;387:894-906. [Crossref] [PubMed]
3. Hora M, Albiges L, Bedke J, Campi R, Capitanio U, Giles RH, Ljungberg B, Marconi L, Klatte T, Volpe A, Abu-Ghanem Y, Dabestani S, Fernández-Pello S, Hofmann F, Kuusk T, Tahbaz R, Powles T, Bex A, Trpkov K. European Association of Urology Guidelines Panel on Renal Cell Carcinoma Update on the New World Health Organization Classification of Kidney Tumours 2022: The Urologist's Point of View. Eur Urol 2023;83:97-100. [Crossref] [PubMed]
4. Majdoub M, Yanagisawa T, Quhal F, Laukhtina E, von Deimling M, Kawada T, Rajwa P, Bianchi A, Pallauf M, Mostafaei H, Chlosta M, Pradere B, Karakiewicz PI, Schmidinger M, Rub R, Shariat SF. Role of clinicopathological variables in predicting recurrence and survival outcomes after surgery for non-metastatic renal cell carcinoma: Systematic review and meta-analysis. Int J Cancer 2024;154:1309-23. [Crossref] [PubMed]
5. Capitanio U, Bedke J, Albiges L, Volpe A, Giles RH, Hora M, Marconi L, Klatte T, Abu-Ghanem Y, Dabestani S, Fernández Pello S, Hofmann F, Kuusk T, Tahbaz R, Powles T, Ljungberg B, Bex A. A Renewal of the TNM Staging System for Patients with Renal Cancer To Comply with Current Decision-making: Proposal from the European Association of Urology Guidelines Panel. Eur Urol 2023;83:3-5. [Crossref] [PubMed]
6. Huang EP, O'Connor JPB, McShane LM, Giger ML, Lambin P, Kinahan PE, Siegel EL, Shankar LK. Criteria for the translation of radiomics into clinically useful tests. Nat Rev Clin Oncol 2023;20:69-82. [Crossref] [PubMed]
7. Ferro M, Musi G, Marchioni M, Maggi M, Veccia A, Del Giudice F, Barone B, Crocetto F, Lasorsa F, Antonelli A, Schips L, Autorino R, Busetto GM, Terracciano D, Lucarelli G, Tataru OS. Radiogenomics in Renal Cancer Management-Current Evidence and Future Prospects. Int J Mol Sci 2023;24:4615. [Crossref] [PubMed]
8. Maddalo M, Bertolotti L, Mazzilli A, Flore AGM, Perotta R, Pagnini F, Ziglioli F, Maestroni U, Martini C, Caruso D, Ghetti C, De Filippo M. Small Renal Masses: Developing a Robust Radiomic Signature. Cancers (Basel) 2023;15:4565. [Crossref] [PubMed]
9. Demirjian NL, Varghese BA, Cen SY, Hwang DH, Aron M, Siddiqui I, Fields BKK, Lei X, Yap FY, Rivas M, Reddy SS, Zahoor H, Liu DH, Desai M, Rhie SK, Gill IS, Duddalwar V. CT-based radiomics stratification of tumor grade and TNM stage of clear cell renal cell carcinoma. Eur Radiol 2022;32:2552-63. [Crossref] [PubMed]
10. Gao Y, Wang X, Zhao X, Zhu C, Li C, Li J, Wu X. Multiphase CT radiomics nomogram for preoperatively predicting the WHO/ISUP nuclear grade of small (< 4 cm) clear cell renal cell carcinoma. BMC Cancer 2023;23:953. [Crossref] [PubMed]
11. Wang J, Huang Z, Zhou J. Radiomics Model for Predicting FOXP3 Expression Level and Survival in Clear Cell Renal Carcinoma. Acad Radiol 2024;31:1447-59. [Crossref] [PubMed]
12. Heller N, Isensee F, Maier-Hein KH, Hou X, Xie C, Li F, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge. Med Image Anal 2021;67:101821. [Crossref] [PubMed]
13. National Comprehensive Cancer Network. NCCN Clinical Practice Guidelines in Oncology: Kidney Cancer. Version 2.2012. Fort Washington, PA: NCCN; 2012.
14. Motzer RJ, Jonasch E, Agarwal N, Beard C, Bhayani S, Bolger GB, et al. Kidney cancer, version 3.2015. J Natl Compr Canc Netw 2015;13:151-9. [Crossref] [PubMed]
15. Zwanenburg A, Vallières M, Abdalah MA, Aerts HJWL, Andrearczyk V, Apte A, et al. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology 2020;295:328-38. [Crossref] [PubMed]
16. Wu J, Li J, Huang B, Dong S, Wu L, Shen X, Zheng Z. Radiomics predicts the prognosis of patients with clear cell renal cell carcinoma by reflecting the tumor heterogeneity and microenvironment. Cancer Imaging 2024;24:124. [Crossref] [PubMed]
17. Cornejo KM, Rice-Stitt T, Wu CL. Updates in Staging and Reporting of Genitourinary Malignancies. Arch Pathol Lab Med 2020;144:305-19. [Crossref] [PubMed]
18. Yang H, Wu K, Liu H, Wu P, Yuan Y, Wang L, Liu Y, Zeng H, Li J, Liu W, Wu S. An automated surgical decision-making framework for partial or radical nephrectomy based on 3D-CT multi-level anatomical features in renal cell carcinoma. Eur Radiol 2023;33:7532-41. [Crossref] [PubMed]
19. Dewi TNK, Perkasa YS, Syaja’Ah KN, Nurmalasari RR, editors. Noninvasive Pathological Staging of Clear Cell Renal Cell Carcinoma using Computed Tomography-based Radiomics Features and Machine Learning. 2023 17th International Conference on Telecommunication Systems, Services, and Applications (TSSA); IEEE; 2023:1-7.
20. Zeng S, Wang XL, Yang H. Radiomics and radiogenomics: extracting more information from medical images for the diagnosis and prognostic prediction of ovarian cancer. Mil Med Res 2024;11:77. [Crossref] [PubMed]
21. He H, Jin Z, Dai J, Wang H, Sun J, Xu D. Computed tomography-based radiomics prediction of CTLA4 expression and prognosis in clear cell renal cell carcinoma. Cancer Med 2023;12:7627-38. [Crossref] [PubMed]
22. Tang X, Pang T, Yan WF, Qian WL, Gong YL, Yang ZG. The Prognostic Value of Radiomics Features Extracted From Computed Tomography in Patients With Localized Clear Cell Renal Cell Carcinoma After Nephrectomy. Front Oncol 2021;11:591502. [Crossref] [PubMed]
23. Zhang T, Ming Y, Xu J, Jin K, Huang C, Duan M, Li K, Liu Y, Lv Y, Zhang J, Huang Z. Radiomics and Ki-67 index predict survival in clear cell renal cell carcinoma. Br J Radiol 2023;96:20230187. [Crossref] [PubMed]
24. Nie P, Yang G, Wang Y, Xu Y, Yan L, Zhang M, Zhao L, Wang N, Zhao X, Li X, Cheng N, Wang Y, Chen C, Wang N, Duan S, Wang X, Wang Z. A CT-based deep learning radiomics nomogram outperforms the existing prognostic models for outcome prediction in clear cell renal cell carcinoma: a multicenter study. Eur Radiol 2023;33:8858-68. [Crossref] [PubMed]
25. Gao R, Qin H, Lin P, Ma C, Li C, Wen R, Huang J, Wan D, Wen D, Liang Y, Huang J, Li X, Wang X, Chen G, He Y, Yang H. Development and Validation of a Radiomic Nomogram for Predicting the Prognosis of Kidney Renal Clear Cell Carcinoma. Front Oncol 2021;11:613668. [Crossref] [PubMed]
26. Raman AG, Fisher D, Yap F, Oberai A, Duddalwar VA. Radiomics and Artificial Intelligence: Renal Cell Carcinoma. Urol Clin North Am 2024;51:35-45. [Crossref] [PubMed]
27. Lu C, Shiradkar R, Liu Z. Integrating pathomics with radiomics and genomics for cancer prognosis: A brief review. Chin J Cancer Res 2021;33:563-73. [Crossref] [PubMed]
28. Brancato V, Cavaliere C, Garbino N, Isgrò F, Salvatore M, Aiello M. The relationship between radiomics and pathomics in Glioblastoma patients: Preliminary results from a cross-scale association study. Front Oncol 2022;12:1005805. [Crossref] [PubMed]
29. Meng X, Shu J, Xia Y, Yang R. A CT-Based Radiomics Approach for the Differential Diagnosis of Sarcomatoid and Clear Cell Renal Cell Carcinoma. Biomed Res Int 2020;2020:7103647. [Crossref] [PubMed]
30. Jiang Y, Li W, Huang C, Tian C, Chen Q, Zeng X, Cao Y, Chen Y, Yang Y, Liu H, Bo Y, Luo C, Li Y, Zhang T, Wang R. A Computed Tomography-Based Radiomics Nomogram to Preoperatively Predict Tumor Necrosis in Patients With Clear Cell Renal Cell Carcinoma. Front Oncol 2020;10:592. [Crossref] [PubMed]
31. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol 2022;19:132-46. [Crossref] [PubMed]
32. Ning Z, Pan W, Chen Y, Xiao Q, Zhang X, Luo J, Wang J, Zhang Y. Integrative analysis of cross-modal features for the prognosis prediction of clear cell renal cell carcinoma. Bioinformatics 2020;36:2888-95. [Crossref] [PubMed]
33. Schulz S, Woerl AC, Jungmann F, Glasner C, Stenzel P, Strobl S, Fernandez A, Wagner DC, Haferkamp A, Mildenberger P, Roth W, Foersch S. Multimodal Deep Learning for Prognosis Prediction in Renal Cancer. Front Oncol 2021;11:788740. [Crossref] [PubMed]
34. Boehm KM, Khosravi P, Vanguri R, Gao J, Shah SP. Harnessing multimodal data integration to advance precision oncology. Nat Rev Cancer 2022;22:114-26. [Crossref] [PubMed]
35. Powles T, Albiges L, Bex A, Comperat E, Grünwald V, Kanesvaran R, Kitamura H, McKay R, Porta C, Procopio G, Schmidinger M, Suarez C, Teoh J, de Velasco G, Young M, Gillessen SESMO Guidelines Committee. Electronic address: clinicalguidelines@esmo. Renal cell carcinoma: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol 2024;35:692-706. [Crossref] [PubMed]
36. Mielczarek Ł, Brodziak A, Sobczuk P, Kawecki M, Cudnoch-Jędrzejewska A, Czarnecka AM. Renal toxicity of targeted therapies for renal cell carcinoma in patients with normal and impaired kidney function. Cancer Chemother Pharmacol 2021;87:723-42. [Crossref] [PubMed]
37. Motzer RJ, Jonasch E, Agarwal N, Alva A, Bagshaw H, Baine M, et al. NCCN Guidelines® Insights: Kidney Cancer, Version 2.2024. J Natl Compr Canc Netw 2024;22:4-16. [Crossref] [PubMed]
38. Dercle L, McGale J, Sun S, Marabelle A, Yeh R, Deutsch E, Mokrane FZ, Farwell M, Ammari S, Schoder H, Zhao B, Schwartz LH. Artificial intelligence and radiomics: fundamentals, applications, and challenges in immunotherapy. J Immunother Cancer 2022;10:e005292. [Crossref] [PubMed]
39. Rundo L, Militello C. Image biomarkers and explainable AI: handcrafted features versus deep learned features. Eur Radiol Exp 2024;8:130. [Crossref] [PubMed]
40. Demircioğlu A. Reproducibility and interpretability in radiomics: a critical assessment. Diagn Interv Radiol 2025;31:321-8. [Crossref] [PubMed]
41. Teng X, Wang Y, Nicol AJ, Ching JCF, Wong EKY, Lam KTC, Zhang J, Lee SW, Cai J. Enhancing the Clinical Utility of Radiomics: Addressing the Challenges of Repeatability and Reproducibility in CT and MRI. Diagnostics (Basel) 2024;14:1835. [Crossref] [PubMed]
42. Peng X, Yang S, Zhou L, Mei Y, Shi L, Zhang R, Shan F, Liu L. Repeatability and Reproducibility of Computed Tomography Radiomics for Pulmonary Nodules: A Multicenter Phantom Study. Invest Radiol 2022;57:242-53. [Crossref] [PubMed]