Preoperative CT Habitat Analysis for Predicting WHO/ISUP Grade in Clear Cell Renal Cell Carcinoma: A Multicenter Study

Preoperative CT Habitat Analysis for Predicting WHO/ISUP Grade in Clear Cell Renal Cell Carcinoma: A Multicenter Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Preoperative CT Habitat Analysis for Predicting WHO/ISUP Grade in Clear Cell Renal Cell Carcinoma: A Multicenter Study Zhe Jin, Siyi Chen, Chen Jin, Ying Ma, Yunshi Liang, Jingxuan Guo, and 11 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9303146/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Purpose To develop and validate a preoperative contrast-enhanced CT–based approach that leverages intratumoral spatial heterogeneity to predict WHO/ISUP grading and prognosis in clear cell renal cell carcinoma (ccRCC). Methods This multicenter retrospective study included 704 patients with pathologically confirmed ccRCC (483 low-grade and 221 high-grade). Patients from SYUCC (n = 308) were randomly split into training and internal validation sets at an 8:2 ratio. Two independent external validation cohorts were used, including GZFPH (n = 106) and two public datasets, KiTS19 (n = 142) and TCGA-KIRC (n = 148). We constructed three models: a conventional whole-tumor radiomics model, an intratumoral habitat model capturing spatial heterogeneity, and a Combined model using score-level fusion of radiomic and habitat scores. Prognostic value was assessed in KiTS19 and TCGA using Kaplan–Meier analysis with log-rank tests based on (i) ground-truth WHO/ISUP grade and (ii) model-derived risk groups. Biological interpretability was explored using differential expression and pathway enrichment analyses, GSVA-based pathway activity mapping to habitat subregions, and immune profiling (IPS and MCPcounter). Results The Combined model achieved consistent performance across cohorts, with AUCs of 0.860 (95% CI: 0.813–0.907) in SYUCC, 0.830 (0.813–0.907) in GZFPH, 0.829 (0.756–0.901) in KiTS19, and 0.750 (0.664–0.830) in TCGA. Although AUC differences between the Combined and Habitat models were not statistically significant, the Combined model showed higher accuracy across all cohorts. Model-derived risk stratification significantly separated overall survival in both KiTS19 (p = 0.017) and TCGA (p = 0.0032). Transcriptomic analyses indicated coherent biological axes involving immune effector and proliferation programs versus differentiation and EMT/TGF-β–related states, with GSVA revealing directionally distinct pathway associations for habitat subregions (S1 vs S3). Immune analyses further supported risk-group differences in antigen presentation and effector immunity components and in inferred immune cell infiltration. Conclusion Intratumoral spatial heterogeneity on preoperative contrast-enhanced CT enables robust prediction of WHO/ISUP grade and prognostic stratification in ccRCC. Clear cell renal cell carcinoma WHO/ISUP grading Computed tomography Radiomics Habitat Analysis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Clear cell renal cell carcinoma (ccRCC) is the most common histological subtype of renal cell carcinoma, accounting for approximately 70% of cases and contributing substantially to kidney cancer–related mortality 1 , 2 . The WHO / International Society of Urological Pathology (ISUP) grading system is a critical prognostic determinant in ccRCC, closely associated with tumor aggressiveness, risk of recurrence and metastasis, and long-term survival 3 , 4 . Compared with low-grade tumors (grades I–II), high-grade tumors (grades III–IV) generally indicate more aggressive tumor biology and poorer clinical outcomes 3 . Accordingly, accurate assessment of WHO / ISUP grade is essential for individualized clinical management, including surgical planning, perioperative risk stratification, and the tailoring of postoperative surveillance and adjuvant treatment strategies 5 . However, reliable preoperative determination of WHO/ISUP grade remains challenging. Percutaneous renal tumour biopsy is a commonly used method for preoperative grading of ccRCC, yet it is invasive and prone to sampling error, particularly in the presence of marked intratumoral heterogeneity, which can lead to grade underestimation or discordant results 6 . These limitations constrain its preoperative utility and underscore the necessity of a noninvasive, reproducible imaging approach for WHO/ISUP grade prediction to support risk-adapted decision-making in ccRCC.Computed tomography (CT) is the most widely used imaging modality for preoperative evaluation of renal masses, playing a central role in tumor localization, characterization of enhancement patterns, and preoperative staging 7 , 8 . With the advancement of radiomics, quantitative image analysis enables the extraction of high-dimensional features that capture tumor morphology, intensity distribution, and textural heterogeneity, thereby enabling noninvasive preoperative prediction of pathological characteristics 9 , 10 . In recent years, several studies have developed CT- or MRI-based radiomics models to predict the nuclear grade of ccRCC 11 – 14 . However, most existing models rely on whole-tumor regions of interest (ROIs) for feature extraction. Such global characterization may obscure regional differences by averaging heterogeneous tumor components, thereby limiting the ability to fully capture intratumoral spatial heterogeneity (ITH) 15 . This spatial heterogeneity may reflect distinct molecular alterations, proliferative activity, and malignant potential across tumor subregions 16 . Moreover, the biological underpinnings of radiomic features remain insufficiently elucidated, which in turn restricts model interpretability and limits robust generalization across institutions 16 . To better characterize intratumoral spatial heterogeneity, recent studies in several solid tumors, including breast cancer, have increasingly explored unsupervised subregional partitioning strategies 17 , 18 . This strategy automatically partitions tumors into subregions with similar imaging characteristics based on voxel-level intensity and texture distributions, and subsequently quantifies spatial organization using ecological diversity or heterogeneity indices. Habitat-based modeling therefore enables a more refined characterization of ITH and has demonstrated potential to enhance predictive performance and interpretability. ccRCC typically exhibits marked spatial heterogeneity on imaging, with substantial variations in enhancement patterns, grayscale distribution, and textural architecture across tumor regions 19 . Such heterogeneity may reflect the complex “ecological architecture” of the tumor and correspond to distinct biological states and malignant potential. However, the role of ITH-driven subregional modeling in predicting WHO/ISUP grade in ccRCC remains insufficiently investigated. In addition, the clinical translation of radiomics models fundamentally depends on their biological interpretability 16 . Emerging evidence from radiogenomic studies suggests that imaging phenotypes may partially reflect underlying molecular characteristics, metabolic reprogramming, angiogenesis, and the immune microenvironment, thereby enabling the establishment of verifiable links between imaging features and tumor biology 20 . However, in the context of ccRCC nuclear grade prediction, most prior studies have focused on constructing discriminative models and reporting performance metrics, with limited investigation into the biological pathways and immunological correlates underlying key radiomic features—particularly those related to intratumoral spatial heterogeneity. Leveraging external cohorts with transcriptomic data to conduct pathway enrichment and immune infiltration analyses is therefore crucial for elucidating the biological basis of ITH-associated imaging phenotypes and strengthening model credibility and translational applicability. Therefore, in this study, we aimed to evaluate the performance and clinical utility of CT-derived intratumoral spatial heterogeneity for preoperative prediction of WHO/ISUP grade and prognostic risk stratification in ccRCC. Furthermore, by leveraging an external cohort with available transcriptomic data, we performed radiogenomic and immune-related analyses to investigate the molecular pathways and immune microenvironment characteristics associated with ITH imaging phenotypes. This integrative approach may provide biological insights into the proposed model and enhance its clinical credibility and translational potential. Method and Materials Patients and study design This multicenter retrospective study was approved by the institutional review boards of all participating hospitals and was conducted in accordance with the Declaration of Helsinki. Due to its retrospective nature, written informed consent for retrospective datasets was waived. Four cohorts were included, comprising two institutional cohorts and two public datasets. The institutional cohorts were obtained from Sun Yat-sen University Cancer Center (SYUCC) and Guangzhou First People’s Hospital (GZFPH). The inclusion criteria were: (1) pathologically confirmed ccRCC after surgery; (2) no concurrent other malignant tumors; (3) underwent radical or partial nephrectomy; and (4) available preoperative abdominal CT. The exclusion criteria were: (1) absence of pathological WHO/ISUP nuclear grade information; (2) lack of contrast-enhanced CT or poor image quality; and (3) incomplete clinical data. In the institutional cohorts, 308 patients from SYUCC and 106 patients from GZFPH were finally included. In addition, two public datasets were included: KiTS19 (Kidney Tumor Segmentation 2019 [KiTS19]; MICCAI 2019 Kidney Tumor Segmentation Challenge; https://github.com/neheller/kits19 ) and TCGA-KIRC (The Cancer Genome Atlas [TCGA]; https://www.cancer.gov/tcga ; TCGA-KIRC). For KiTS19, 210 patients were initially identified; after excluding 67 patients with non-ccRCC histology and 1 patient with poor image quality, 142 patients were included. For TCGA-KIRC, patients without contrast-enhanced CT, with poor image quality, or with incomplete clinical information were excluded, resulting in 148 ccRCC patients. The detailed recruitment flowcharts are shown in Fig. 1 . In total, 704 patients were included. The SYUCC cohort was randomly divided into training and internal validation sets at a 8:2 ratio for model development. The GZFPH, KiTS19, and TCGA-KIRC cohorts served as independent external validation cohorts. The study was conducted in four phases: Phase 1: Model development; Phase 2: Multi-center validation; Phase 3: Prognostic prediction analysis; Phase 4: Biological interpretability analysis. Pathological assessment and clinical outcomes Pathological data were retrieved from postoperative pathology reports in the hospital information system. Pathological diagnosis and grading were assessed by experienced senior pathologists according to the 2016 WHO classification criteria 21 . For analyses, WHO/ISUP grades I–II were categorized as low grade, and grades III–IV as high grade. Clinical and pathological variables collected included age, sex, tumor laterality, tumor size, clinical T stage, and ISUP grade. Follow-up information for the KiTS19 and TCGA-KIRC cohorts has been previously reported in the corresponding public dataset documentation and related studies 22 , 23 . Overall survival (OS) data were extracted to evaluate the prognostic stratification performance of the proposed model. All 142 patients included from KiTS19 had available OS data, and 128 patients in the TCGA-KIRC cohort had available OS data. OS was defined as the time from initial surgery to death from any cause, or to the last follow-up for censored patients. CT imaging acquisition All patients underwent preoperative contrast-enhanced abdominal CT examinations within 2 weeks prior to surgery. CT images were acquired at multiple institutions using multi-detector CT scanners, including Siemens, Toshiba, GE, and Philips systems. Four-phase CT scans were routinely performed, including unenhanced phase, corticomedullary phase, nephrographic phase, and excretory phase. Nonionic iodinated contrast medium was intravenously administered using a power injector at a dose of 1–1.5 mL/kg (maximum 150 mL), with an injection rate of 2.5–3.5 mL/s. For model development, the corticomedullary phase images were selected for subsequent analysis, based on previous evidence suggesting that this phase provides optimal tumor–renal parenchyma contrast and better differentiation between high-grade and low-grade ccRCC 22 . Detailed CT acquisition parameters for each institution are summarized in Supplementary Material 1 . Image preprocessing and subregion segmentation The overall workflow is summarized in Fig. 2 , and detailed procedures are described below. Tumor subregions were generated using a two-stage unsupervised strategy. Two experienced radiologists (T.W.J. and G.Y., with 10 and 18 years of experience in abdominal imaging, respectively) manually delineated the three-dimensional tumor volume of interest (VOI) on contrast-enhanced CT using ITK-SNAP (v3.8.0). To enhance cross-center consistency, CT images were first resampled to isotropic voxel spacing (1×1×1 mm³). Within each tumor VOI, voxel intensities were clipped using percentile-based truncation to mitigate scanner-dependent extreme values. Robust intensity normalization was subsequently applied to reduce the influence of residual outliers, followed by z-score standardization to harmonize feature scales across patients. In Stage 1, mask-constrained three-dimensional supervoxels were generated using the SLIC algorithm implemented in SimpleITK. Supervoxel size was set to 8×8×8 voxels in isotropic space, and the compactness parameter was fixed at 0.4 to balance spatial regularity and boundary adherence, thereby producing spatially contiguous and anatomically coherent initial units. In Stage 2, a predefined 20-dimensional feature vector was extracted from each supervoxel, primarily consisting of first-order intensity statistics. During model training, these feature vectors were standardized and mapped into a global prototype space constructed exclusively from training-set samples. Supervoxels were assigned to prototype categories via nearest-neighbor matching. During inference, weighted K-means clustering (weighted by supervoxel volume) was performed within each case to obtain N subregions. The optimal number of subregions was determined using bayesian information criterion analysis on the training set, which indicated N = 3 as the most parsimonious and stable solution. Case-specific clusters were then aligned with the global prototype labels using a one-to-one greedy matching strategy to ensure consistent, non-overlapping subregion labeling across patients. This procedure yielded refined tumor habitats characterized by improved internal homogeneity and cross-patient label consistency. Model development We first applied minimum redundancy maximum relevance to pre-screen the initial imaging features and reduce redundancy in the high-dimensional feature space. For both intratumoral habitat heterogeneity features and conventional radiomics features, feature repeatability was evaluated prior to subsequent selection. Specifically, five pseudo-masks were generated to simulate contour perturbations of manual delineation, and features were repeatedly extracted from each pseudo-mask to calculate intraclass correlation coefficients (ICCs). Features were considered repeatable and retained for downstream analyses if ICC > 0.90 for habitat features and ICC > 0.75 for radiomics features (a stricter ICC threshold was used for habitat features because they rely more on subregion partitioning and are more sensitive to boundary perturbations). Next, low-variance filtering was performed to remove uninformative features, followed by Pearson correlation analysis to eliminate highly correlated features (|r| > 0.75) and mitigate multicollinearity. Candidate features significantly associated with the endpoint were then identified using an independent-samples t test or the Mann–Whitney U test, as appropriate (P < 0.05). Based on the above statistical filtering steps, a second-stage feature selection was conducted using a genetic algorithm combined with a support vector machine (SVM) to obtain a more robust and generalizable feature subset. The training set was split into a sub-training set and an internal validation set. Each GA individual represented a fixed-length feature subset, and the fitness function was defined by the classification performance of an SVM model (AUC and accuracy) on both the sub-training and internal validation sets. The optimal feature subset was selected and saved. Finally, 20 habitat features and 20 radiomics features were determined for model construction ( Supplementary Material 2 ). Using the final feature sets, three predictive models were developed: (1) a Radiomics model, (2) a Habitat model, and (3) a Combined model. Both the Radiomics and Habitat models were built using logistic regression (LR) to generate their respective rad-scores. For the Combined model, score-level fusion was implemented by feeding the two rad-scores into a second-stage logistic regression model to derive a combined rad-score for final classification. All model development was implemented in Python (scikit-learn v1.0.2), and model selection and performance evaluation were conducted using the training and internal validation sets. Model assessment All models were developed and evaluated using Python (v3.7.7) with scikit-learn (v1.0.2). Hyperparameter tuning was performed exclusively within the training set using grid search combined with five-fold cross-validation, and the model with the best mean cross-validated performance was selected. The finalized models were subsequently evaluated on the internal validation set and three independent external cohorts (GZFPH, KiTS19, and TCGA-KIRC). To assess the incremental value of different feature sets, we constructed and compared a Radiomics model based on conventional radiomics features, a Habitat model derived from intratumoral subregion (habitat) characteristics, and a Combined model integrating both feature types. Discrimination for WHO/ISUP grade (high vs low) was evaluated using receiver operating characteristic (ROC) curves and quantified by the area under the ROC curve (AUC). Sensitivity, specificity and accuracy were c alculated at a prespecified threshold. AUCs were compared using the DeLong test, and 95% confidence intervals were estimated via bootstrap resampling. To enhance model interpretability, intratumoral habitat segmentation maps were spatially visualized. Subregion labels were projected back onto the original CT images and overlaid using color-coded representations to illustrate spatial heterogeneity patterns within tumors. Representative cases with true-positive, true-negative, false-positive, and false-negative predictions were presented to qualitatively demonstrate the relationship between spatial subregion distribution and model predictions. Biological interpretability analysis To explore the biological underpinnings associated with the model-derived risk stratification, we performed a multi-level biological interpretability analysis at the transcriptomic level using TCGA ccRCC cohort. Gene expression data and corresponding clinical information were obtained from the TCGA database. Patients were stratified into high- and low-grade groups according to the model-derived risk score. Differential pathway enrichment analysis between groups was conducted using Gene Set Enrichment Analysis (GSEA). The analysis was implemented with the clusterProfiler, msigdbr, and GSEAbase R packages based on the c2.cp and Reactome gene sets from the Molecular Signatures Database. Enriched pathways were identified using the criteria of nominal P 1. To further quantify pathway-level activity in individual samples, Gene Set Variation Analysis (GSVA) was performed using the GSVA R package. The limma package was subsequently applied to assess differential pathway activity between groups. Gene sets were considered significantly different with P 0.1. To evaluate the association between deep learning-derived imaging features and biological pathways, Spearman correlation analysis was conducted between model features and GSVA-derived pathway scores. Correlations with P < 0.05 were considered statistically significant. Tumor microenvironment (TME) cellular composition was estimated using single-sample Gene Set Enrichment Analysis (ssGSEA) implemented in the GSVA package. Cell-type-specific enrichment scores were calculated for each sample, and differences in immune cell abundance between risk groups were assessed using nonparametric statistical tests. Survival analysis Survival analyses were conducted in the KiTS19 and TCGA cohorts. OS was estimated using the Kaplan–Meier method and compared using the log-rank test. Within each cohort, OS was analyzed under two stratification schemes: (i) the ground-truth WHO/ISUP grade (true low-grade vs true high-grade) and (ii) the model-derived risk groups. For the model-based stratification, patients were dichotomized into low- and high-risk groups according to the cohort-specific median value of the Combined model–derived rad-score, and Kaplan-Meier curves were generated accordingly. All statistical tests were two-sided, and P < 0.05 was considered statistically significant. Statistical analyses were performed using R (version 3.4.1), SPSS (version 23.0; IBM, Armonk, NY, USA), and Python (version 3.12). Results Clinicopathologica characteristics A total of 704 patients with pathologically confirmed ccRCC were included, comprising 483 low-grade (WHO/ISUP I–II) and 221 high-grade (III–IV) tumors. The SYUCC cohort (n = 308) served as the development set, and the GZFPH (n = 106), KiTS19 (n = 142), and TCGA-KIRC (n = 148) cohorts were used for external validation. Baseline clinicopathologic characteristics across the four cohorts are summarized in Table 1 . Table 1 Baseline clinicopathologic characteristics of patients with ccRCC in the four cohorts. Characteristics SYUCC cohort (N = 308) GZFPH cohort (N = 106) KiTS19 cohort (N = 142) TCGA cohort (N = 148) Age, years 52.0 ± 12.4 59 ± 12.8 59.3 ± 12.6 60.4 ± 12.7 Sex, n (%) Male 219(71.1) 71(67.0) 92(64.8) 89(60.1) Female 89(28.9) 35(33.0) 50(35.2) 59(39.9) Side, n (%) Right 147(47.7) 49(46.2) 70(49.3) 79(53.4) Left 161(52.3) 57(53.8) 72(50.7) 69(46.6) Maximum tumour diameter, cm 4.2 ± 2.1 5.2 ± 2.5 4.8 ± 3.0 6.2 ± 3.0 T staging, n (%) T1 267(86.7) 81(76.4) 100(70.4) 87(58.8) T2 20(6.5) 10(9.4) 7(4.9) 14(9.5) T3 21(6.8) 13(12.3) 34(23.9) 45(30.4) T4 0(0.0) 2(1.9) 1(0.7) 2(1.3) WHO/ISUP grade, n (%) I 25(8.1) 23(21.7) 19(13.4) 1(0.7) II 205(66.6) 65(61.3) 82(57.7) 63(42.6) III 68(22.1) 12(11.3) 31(21.8) 61(41.2) IV 10(3.2) 6(5.7) 10(7.0) 23(15.5) Data were presented as number of patients, with the exception of age and maximum tumour diamete at baseline (mean ± SD). Abbreviations: ccRCC = clear cell renal cell carcinoma; ISUP = International Society of Urological Pathology; SD= standard deviation. Across cohorts, the mean age ranged from 52.0 to 60.4 years. Males represented the majority of patients in each cohort, and tumor laterality was generally balanced between the right and left kidneys. The mean maximum tumor diameter varied across cohorts, ranging from 4.2 cm in the SYUCC cohort to 6.2 cm in the TCGA cohort. Most tumors were stage T1 in all cohorts, although the proportion of T1 disease was lower in the TCGA and KiTS19 cohorts than in the SYUCC and GZFPH cohorts. Predictive performance of the models for WHO/ISUP grading The predictive performance of the Radiomics, Habitat, and Combined models for WHO/ISUP grade classification is summarized in Table 2 and Fig. 3 a–d. Overall, the Combined model showed the highest performance across cohorts, with AUCs of 0.860 (95% CI: 0.813–0.907) in the SYUCC cohort, 0.830 (0.813–0.907) in the GZFPH cohort, 0.829 (0.756–0.901) in the KiTS19 cohort, and 0.750 (0.664–0.830) in the TCGA cohort. Although the AUC differences between the Combined and Habitat models were not statistically significant across cohorts (P > 0.05), the Combined model demonstrated higher overall accuracy in all datasets (0.805, 0.764, 0.796, and 0.730, respectively), together with favorable sensitivity and specificity trade-offs in several cohorts. The Habitat model also demonstrated strong discriminative performance, with consistently higher AUCs than the Radiomics model across all cohorts (SYUCC: 0.830 vs 0.800; GZFPH: 0.814 vs 0.797; KiTS19: 0.812 vs 0.792; TCGA: 0.750 vs 0.610). Notably, in the TCGA cohort, both the Habitat and Combined models improved discrimination compared with the Radiomics model (AUC 0.750 vs 0.610). Table 2 Performance of radiomics, habitat, and combined models for WHO/ISUP grade prediction across four cohorts. Cohorts Model AUC (95% CI) Accuracy Sensitivity Specificity P-value SYUCC cohort Radiomics 0.800 (0.740–0.860) 0.779 0.770 0.783 0.013 Habitat 0.830 (0.775–0.885) 0.795 0.805 0.792 0.072 Combined 0.860 (0.813–0.907) 0.805 0.816 0.801 Ref GZFPH cohort Radiomics 0.797 (0.707–0.877) 0.736 0.737 0.736 0.146 Habitat 0.814 (0.735–0.888) 0.708 0.947 0.655 0.575 Combined 0.830 (0.813–0.907) 0.764 0.842 0.747 Ref KiTS19 cohort Radiomics 0.792 (0.707–0.877) 0.732 0.756 0.723 0.060 Habitat 0.812 (0.735–0.888) 0.725 0.878 0.663 0.593 Combined 0.829 (0.756–0.901) 0.796 0.780 0.802 Ref TCGA cohort Radiomics 0.610 (0.519–0.701) 0.608 0.595 0.625 0.006 Habitat 0.750 (0.668–0.834) 0.716 0.714 0.719 0.496 Combined 0.750 (0.664–0.830) 0.730 0.774 0.672 Ref Abbreviations: AUC = area under the receiver operating characteristics curve; 95% CI = 95% confidence interval. Survival prediction Kaplan-Meier analyses were performed in the KiTS19 and TCGA cohorts to evaluate OS stratification by (i) the ground-truth WHO/ISUP grade and (ii) the model-derived risk groups (Fig. 4 ). In the KiTS19 cohort, OS differed significantly between true low-grade and true high-grade tumors (log-rank p = 0.012), and a significant separation was also observed between the model-defined low-risk and high-risk groups (log-rank p = 0.017). In the TCGA cohort, OS stratification by the ground-truth WHO/ISUP grade did not reach statistical significance (log-rank p = 0.065), whereas the model-derived risk stratification showed significant separation between the low-risk and high-risk groups (log-rank p = 0.0032), with the high-risk group exhibiting consistently lower survival probability over follow-up. Gene expression and pathway enrichment analysis To characterize transcriptomic differences associated with the Combined model derived risk stratification, we compared differential expression profiles between the high risk and low risk groups. The volcano plot in Fig. 5 a summarizes the differentially expressed genes (DEGs) between the two groups, and Fig. 5 b highlights representative genes upregulated in the high risk group, including IL6, TREM1, WNT10B, COL7A1, and SCO2 (ranked by log2 fold change). Functionally, IL6 and TREM1 are linked to inflammatory and innate immune signaling, COL7A1 is related to extracellular matrix components, WNT10B is involved in WNT associated transcriptional regulation, and SCO2 is associated with mitochondrial energy metabolism. At the gene set level, GSEA based on the MSigDB C2 (curated gene sets) collection showed that gene sets significantly upregulated in the high risk group were primarily enriched for immune effector and proliferation related processes (Fig. 5 c), exemplified by the natural killer cell related gene set (CURSONS_NATURAL_KILLER_CELLS) and a cancer proliferation and grade related gene set (CANCER_PROLIFERATION_CLUSTER). In contrast, significantly downregulated gene sets were mainly related to epithelial differentiation and EMT and TGF-β associated transcriptional programs (Fig. 5 d), including the epithelial differentiation module (BOSCO_EPITHELIAL_DIFFERENTIATION_MODULE) and a TGF-β mediated EMT gene set (FOROUTAN_TGFB_EMT_DN). Furthermore, we quantified pathway activities using GSVA and assessed correlations between habitat features and pathway scores (Fig. 5 e; red indicates positive correlations and blue indicates negative correlations). Overall, habitat features were significantly associated with immune effector related pathway modules, represented by antigen presentation and cross presentation and TCR associated signaling, and with cell cycle and DNA replication and proliferation modules, including S phase, G2 and M checkpoints, and DNA replication. At the subregion level, S1 related habitat features showed an overall positive correlation with the immune effector module and the cell cycle and replication module, whereas S3 related features more frequently exhibited correlations in the opposite direction (Fig. 5 e). Immune microenvironment Building on the transcriptomic analyses that revealed immune-related differences between the model-derived risk groups, we further characterized the immune microenvironment using IPS components and MCPcounter-based deconvolution (Fig. 6 ). In the IPS analysis, all four components showed significant between-group differences. Specifically, the high-risk group exhibited higher MHC_IPS (an antigen-presentation-related component; p = 0.0009) and higher EC_IPS (an effector-cell-related component; p = 0.0094) (Fig. 6 a). In addition, SC_IPS (a suppressor-cell-related component; p = 0.0445) and CP_IPS (a checkpoint-related component; p = 0.0002) also differed significantly between the two groups (Fig. 6 a). Consistent with these score-level differences, MCPcounter deconvolution identified significant group-level differences in inferred immune cell infiltration (Fig. 6 b). Compared with the low-risk group, the high-risk group showed higher estimated abundance of cytotoxic lymphocytes (p = 0.0077), neutrophils (p = 0.0132), NK cells (p = 0.0293), and CD8 + T cells (p = 0.0283). Other MCPcounter signatures did not show significant differences. Collectively, these findings indicate that the model-derived risk stratification is associated with distinct immune-related profiles at both the IPS score level and the cell-composition level. Model visualization To provide an intuitive visualization of intratumoral spatial heterogeneity patterns associated with WHO/ISUP grade prediction, we present representative cases with habitat subregion maps overlaid on the tumor ROI and corresponding 3D reconstructions (Fig. 7 ). Patient A (predicted high grade) and Patient B (predicted low grade) exhibited distinct intratumoral habitat configurations. In Patient A, S1 (red) was visually more prominent across both 2D slices and the 3D reconstruction, whereas Patient B showed a more prominent S3 (blue). Overall, these qualitative examples illustrate that model predictions align with distinct S1/S3 spatial heterogeneity patterns and are in line with the pathway-activity mapping results (Fig. 5 e). Discussion In this study, we proposed and validated a preoperative CT-based framework for predicting WHO/ISUP grade in ccRCC. Unlike conventional modeling strategies that treat tumors as homogeneous entities, our approach incorporates intratumoral habitat subregions to characterize spatial heterogeneity. The resulting combined model demonstrated stable discriminative performance across multicenter datasets and two public cohorts, with AUCs of at least 0.750 in all cohorts. At the molecular level, the model-derived risk stratification corresponded to transcriptomic differences mainly involving immune effector activity, cell proliferation, and epithelial differentiation/EMT-related pathways. Further analyses indicated that habitat subregion features were associated with these pathway modules, thereby linking “spatial subregional phenotypes” with “molecular pathway themes” and enabling intuitive visualization of subregional tumor architecture. In addition, stratification based on the model-derived risk score effectively separated overall survival in both the KiTS19 and TCGA cohorts, suggesting that this imaging phenotype not only reflects pathological grading information but also captures prognostically relevant cues of biological aggressiveness. In this study, we introduced a habitat-based framework for feature representation to explicitly quantify intratumoral spatial heterogeneity, rather than modeling tumors solely with whole-lesion averaged phenotypes. This design allows imaging signatures to better reflect the underlying structural and functional differences within the tumor. Across the multicenter cohorts, the Habitat model generally outperformed the Radiomics model, and its AUC did not differ significantly from that of the Combined model. This suggests that subregional heterogeneity alone provides strong discriminative signals for WHO/ISUP grade prediction. Meanwhile, the Combined model consistently achieved higher accuracy in most cohorts, indicating a decision-level gain: building upon the core heterogeneity information captured by habitat features, integrating complementary representations may further stabilize classification and better align with the clinical need for correct grading. Notably, the models maintained relatively consistent discriminative performance across different datasets, supporting their robustness and potential generalizability across heterogeneous multicenter cohorts. Prior CT-based radiomics and deep learning studies have reported encouraging performance for predicting ccRCC WHO/ISUP grade 22 , 24 – 27 . However, many approaches still rely predominantly on whole-tumor features, which may obscure critical intratumoral heterogeneity 22 , 24 – 27 . In addition, some earlier studies were limited by insufficient external validation or small test cohorts, raising concerns regarding generalizability 24 , 25 . In contrast, subregional strategies have been more frequently explored in ccRCC tasks related to tumor aggressiveness, with evidence suggesting that partitioning tumors into distinct subregions can better characterize spatial heterogeneity and, in some settings, improve performance over whole-tumor analysis 28 , 29 . Our habitat framework follows this line of reasoning and further applies it to WHO/ISUP grading, a clinically risk-relevant pathologic stratification metric, enabling heterogeneous imaging phenotypes to contribute to grade prediction in a reproducible manner and to be validated across independent cohorts. Beyond grading prediction performance, we further evaluated the association between the model-derived risk stratification and OS in the KiTS19 and TCGA cohorts, and compared it with OS stratification based on the ground-truth WHO/ISUP grade. In the KiTS19 cohort, the ground-truth WHO/ISUP grade significantly separated OS (log-rank p = 0.012), supporting its prognostic stratification value in this cohort. In contrast, OS stratification by the ground-truth grade in the TCGA cohort showed only a borderline difference (p = 0.065). Notably, despite the variability in prognostic separation by ground-truth grading across cohorts, the model-derived risk groups consistently discriminated OS in both datasets (KiTS19: p = 0.017; TCGA: p = 0.0032), demonstrating a more uniform prognostic stratification trend 30 . These findings suggest that contrast-enhanced CT–based imaging phenotypes, particularly those capturing spatial heterogeneity, can provide prognostic support in external public cohorts and may offer preoperative survival-related risk cues, thereby serving as a complementary imaging-derived source of information alongside the WHO/ISUP grading system 31 , 32 . To further elucidate the tumor biological implications underlying this imaging-based stratification, we interpreted the model-derived risk groups from a transcriptomic perspective. Overall, the high-risk group exhibited relatively clear thematic axes. First, programs related to immune effector function and inflammation were more active, exemplified by natural killer cell–related pathways and FOXP3-associated immune regulatory programs, accompanied by enhanced cell cycle and proliferation signals 28 , 33 . Second, cell-state programs related to epithelial differentiation and EMT/TGF-β showed a coordinated downregulation, including the epithelial differentiation module, ZEB1 target gene sets, and multiple TGF-β–mediated EMT pathways 28 , 34 . At the DEG level, IL6 and TREM1 point to an inflammatory and innate immune axis, COL7A1 is more closely linked to extracellular matrix and stromal phenotypes, WNT10B relates to cell-state regulation, and SCO2 suggests metabolic adaptation; together, these representative genes converge on the themes above, indicating that imaging features can capture molecular bases associated with tumor biological behavior. Building on these findings, we further performed immune microenvironment analyses to provide cross-level validation of the “immune effector” axis. IPS showed that the high-risk group had higher antigen presentation–related and effector immunity–related components, while immune regulatory components also differed between groups. Consistently, MCPcounter inferred higher infiltration of cytotoxic lymphocytes, NK cells, and CD8 T cells in the high-risk group, together with differences in inflammation-related cellular components, suggesting that this stratification corresponds to a more complex immune ecosystem. Notably, prior imaging phenotype studies in ccRCC prognostic prediction have also reported concordance between imaging-derived stratification and immune infiltration composition, as well as immune-related transcriptional programs such as antigen presentation and interferon signaling, further supporting the potential of imaging phenotypes as indirect readouts of tumor immune status 35 . Collectively, these results provide a biologically coherent interpretive framework for imaging-based stratification of WHO/ISUP grading. To further evaluate the association between habitat features and pathway activity, we applied GSVA to quantify key biological processes and examined their relationships with habitat-derived spatial heterogeneity features. The GSVA results were concordant with the “immune effector and cell proliferation” axis highlighted by GSEA, and further mapped transcriptomic differences onto imaging-defined spatial phenotypes 35 . Notably, habitat features related to S1 (Sub-region 1) showed positive correlations with immune effector modules and cell cycle/replication modules, whereas S3 (Sub-region 3)–related features more frequently exhibited negative correlations. This pattern suggests that distinct spatial subregions may correspond to different spectra of biological activity. Based on this “pathway activity and spatial subregion” mapping, we next used representative case visualizations to illustrate how the spatial configurations of S1 and S3 align with model outputs, thereby providing a more intuitive depiction of the spatial heterogeneity captured by the habitat framework. This study has several limitations. First, although we included multicenter cohorts and incorporated KiTS19 and TCGA for external validation, unavoidable heterogeneity exists across institutions and public databases in CT acquisition protocols, contrast-enhancement phases, reconstruction parameters, and clinical management pathways. Such variability may affect feature extraction and model stability. We mitigated this issue through standardized preprocessing and repeatability-based feature filtering; however, further validation in larger, prospective cohorts with more standardized multiparametric imaging protocols is still warranted. Second, WHO/ISUP grade was used as the primary endpoint, but differences in grade distribution and baseline characteristics across cohorts, together with the relatively limited clinical variables available in public datasets, may introduce potential confounding and influence both model performance and interpretability analyses. Third, prognostic evaluation was primarily based on Kaplan–Meier curves and log-rank tests, and multivariable Cox regression analyses were not performed to adjust for tumor stage, treatment, and other key clinical factors. Therefore, the independent prognostic value of the model-derived risk stratification requires further assessment in cohorts with more complete clinical information. Conclusions This study proposes and validates a contrast-enhanced CT–based habitat framework that captures intratumoral spatial heterogeneity and integrates score-level fusion to enable preoperative prediction of WHO/ISUP grade in clear cell renal cell carcinoma. The framework demonstrated stable predictive performance across multicenter cohorts as well as the public KiTS19 and TCGA cohorts. The model-derived risk score further enabled prognostic stratification of patients with ccRCC. Biological interpretability analyses showed that the risk groups were consistently associated with immune effector activity, proliferation programs, and cell-state–related processes, with clear directionality at the subregion level, thereby providing a biologically coherent interpretive framework for imaging-based stratification. Overall, our findings support spatial heterogeneity–aware imaging phenotypes as a potential noninvasive complement for ccRCC grading and risk assessment, and lay the groundwork for future prospective validation and deeper mechanistic investigations. Abbreviations ccRCC Clear cell renal cell carcinoma ISUP International Society of Urological Pathology CT Computed tomography ROI Region of interest ITH Intratumoral spatial heterogeneity OS Overall survival ICC Intraclass correlation coefficient ROC Receiver operating characteristic AUC Area under the ROC curve GSEA Gene set enrichment analysis DEG Differentially expressed gene Declarations Acknowledgements Not applicable. Author Contributions Zhe Jin, Siyi Chen, and Chen Jin contributed to methodology, investigation, formal analysis, and manuscript drafting. Zhe Jin also contributed to conceptualization, data curation, and visualization. Chen Jin was additionally responsible for software development and validation. Ying Ma, Yunshi Liang, Jingxuan Guo, Luyi Chen, Yingwen Liu, Xusheng Lin, Chuyi Huang, and Zhidan Zhong contributed to data curation and investigation, with Yunshi Liang, Yongxin Chen, Yuan Guo, and Wenjie Tang also contributing resources and/or validation. Weifeng Liu, Li Tian, and Xinqing Jiang contributed to study conception and design, supervision, project administration, and critical revision of the manuscript. Weifeng Liu and Xinqing Jiang also contributed to funding acquisition. All authors contributed to manuscript preparation, read, and approved the final manuscript. Funding This study was supported by the National Natural Science Foundation of China (No. 82302314); Basic and Applied Basic Research Foundation of Guangdong Province (Nos. 2022A1515110792, 2023A1515220097, 2024A1515010653); Science and technology Projects in Guangzhou (Nos. 2024A03J1030, 2025A03J4162, 2025A03J4163); Guangdong Medical Research Fund (No. 202405300148274416); Clinical Research Hongmian Project of Guangzhou First People's Hospital (No. HM2025062) . Data sharing statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Ethics approval and consent to participate This multicenter retrospective study was conducted in accordance with the Declaration of Helsinki. The institutional review boards of Guangzhou First People’s Hospital and Sun Yat-sen University Cancer Center approved the study and waived the requirement for written informed consent due to its retrospective nature. Consent for publication Not applicable. Conflict of Interest All authors declare no competing interests. References Siegel RL, Kratzer TB, Wagle NS, Sung H, Jemal A. Cancer statistics, 2026. CA Cancer J Clin. 2026;76(1):e70043. 10.3322/caac.70043 . From NLM Medline. Bukavina L, Bensalah K, Bray F, Carlo M, Challacombe B, Karam JA, Kassouf W, Mitchell T, Montironi R, O'Brien T, et al. Epidemiology of Renal Cell Carcinoma: 2022 Update. Eur Urol. 2022;82(5):529–42. 10.1016/j.eururo.2022.08.019 . From NLM Medline. Dagher J, Delahunt B, Rioux-Leclercq N, Egevad L, Srigley JR, Coughlin G, Dunglinson N, Gianduzzo T, Kua B, Malone G, et al. Clear cell renal cell carcinoma: validation of World Health Organization/International Society of Urological Pathology grading. Histopathology. 2017;71(6):918–25. 10.1111/his.13311 . From NLM Medline. Mohanty SK, Lobo A, Cheng L. The 2022 revision of the World Health Organization classification of tumors of the urinary system and male genital organs: advances and challenges. Hum Pathol. 2023;136:123–43. 10.1016/j.humpath.2022.08.006 . From NLM Medline. Stewart GD, Klatte T, Cosmai L, Bex A, Lamb BW, Moch H, Sala E, Siva S, Porta C, Gallieni M. The multispeciality approach to the management of localised kidney cancer. Lancet. 2022;400(10351):523–34. 10.1016/S0140-6736(22)01059-5 . From NLM Medline. Harris CR, Whitson JM, Meng MV. Under-grading of < 4 cm renal masses on renal biopsy. BJU Int. 2012;110(6):794–7. 10.1111/j.1464-410X . .2012.10944.x From NLM Medline. Li S, Zhou Z, Gao M, Liao Z, He K, Qu W, Li J, Kamel IR, Chu Q, Zhang Q, et al. Incremental value of automatically segmented perirenal adipose tissue for pathological grading of clear cell renal cell carcinoma: a multicenter cohort study. Int J Surg. 2024;110(7):4221–30. 10.1097/JS9.0000000000001358 . From NLM Medline. Shinagare AB, Davenport MS, Park H, Pedrosa I, Remer EM, Chandarana H, Doshi AM, Schieda N, Smith AD, Vikram R, et al. Lexicon for renal mass terms at CT and MRI: a consensus of the society of abdominal radiology disease-focused panel on renal cell carcinoma. Abdom Radiol (NY). 2021;46(2):703–22. 10.1007/s00261-020-02644-x . From NLM Medline. Gillies RJ, Kinahan PE, Hricak H, Radiomics. Images Are More than Pictures, They Are Data. Radiology. 2016;278(2):563–77. 10.1148/radiol.2015151169 . From NLM Medline. Guiot J, Vaidyanathan A, Deprez L, Zerka F, Danthine D, Frix AN, Lambin P, Bottari F, Tsoutzidis N, Miraglio B, et al. A review in radiomics: Making personalized medicine a reality via routine imaging. Med Res Rev. 2022;42(1):426–40. 10.1002/med.21846 . From NLM Medline. 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(1). 10.1186/s12885-023-11454-5 . Jiao P, Wang B, Ni X, Lu Y, Yang R, Liu Y, Wang J, Mei H, Liu X, Weng X, et al. CT Urography-Based Radiomics to Predict ISUP Grading of Clear Cell Renal Cell Carcinoma. J Cancer. 2025;16(4):1118–26. 10.7150/jca.105173 . Wu C, Xi Y, Hu J, Li G, Wang X, Jiao X, Ding Z, Sun W. A predictive model for WHO/ISUP pathologic grading of renal clear cell carcinoma based on CT radiomics: a multicenter study. BMC Nephrol. 2025;26(1). 10.1186/s12882-025-04268-z . Zhou H, Mao H, Dong D, Fang M, Gu D, Liu X, Xu M, Yang S, Zou J, Yin R, et al. Development and External Validation of Radiomics Approach for Nuclear Grading in Clear Cell Renal Cell Carcinoma. Ann Surg Oncol. 2020;27(10):4057–65. 10.1245/s10434-020-08255-6 . From NLM Medline. Zhang X, Su GH, Chen Y, Gu YJ, You C. Decoding Intratumoral Heterogeneity: Clinical Potential of Habitat Imaging based on Radiomics. Radiology. 2023;309(3):e232047. 10.1148/radiol.232047 . From NLM Medline. Tomaszewski MR, Gillies RJ. The Biological Meaning of Radiomic Features. Radiology. 2021;298(3):505–16. 10.1148/radiol.2021202553 . From NLM Medline. Shi Z, Huang X, Cheng Z, Xu Z, Lin H, Liu C, Chen X, Liu C, Liang C, Lu C, et al. MRI-based Quantification of Intratumoral Heterogeneity for Predicting Treatment Response to Neoadjuvant Chemotherapy in Breast Cancer. Radiology. 2023;308(1):e222830. 10.1148/radiol.222830 . From NLM Medline. Tang W, Jin C, Kong Q, Liu C, Chen S, Ding S, Liu B, Feng Z, Li Y, Dai Y, et al. Development and validation of an MRI spatiotemporal interaction model for early noninvasive prediction of neoadjuvant chemotherapy response in breast cancer: a multicentre study. EClinicalMedicine. 2025;85:103298. 10.1016/j.eclinm.2025.103298 . From NLM PubMed-not-MEDLINE. Gao R, Pang J, Lin P, Wen R, Wen D, Liang Y, Ma Z, Liang L, He Y, Yang H. Identification of clear cell renal cell carcinoma subtypes by integrating radiomics and transcriptomics. Heliyon. 2024;10(11):e31816. 10.1016/j.heliyon.2024.e31816 . From NLM PubMed-not-MEDLINE. Hartmann K, Sadee CY, Satwah I, Carrillo-Perez F, Gevaert O. Imaging genomics: data fusion in uncovering disease heritability. Trends Mol Med. 2023;29(2):141–51. From NLM Medline. Moch H, Cubilla AL, Humphrey PA, Reuter VE, Ulbright TM. The 2016 WHO Classification of Tumours of the Urinary System and Male Genital Organs-Part A: Renal, Penile, and Testicular Tumours. Eur Urol. 2016;70(1):93–105. 10.1016/j.eururo.2016.02.029 . From NLM Medline. Zheng Q, Mei H, Weng X, Yang R, Jiao P, Ni X, Yang X, Wu J, Fan J, Yuan J, et al. Artificial intelligence-based multimodal prediction for nuclear grading status and prognosis of clear cell renal cell carcinoma: a multicenter cohort study. Int J Surg. 2025;111(6):3722–30. 10.1097/JS9.0000000000002368 . From NLM Medline. Heller N, Sathianathen N, Kalapara A, Walczak E, Moore K, Kaluzniak H, Rosenberg J, Blake P, Rengel Z, Oestreich M. The KiTS19 Challenge Data: 300 Kidney Tumor Cases with Clinical Context, CT Semantic Segmentations, and Surgical Outcomes. arXiv e-prints 2019. Luo S, Wei R, Lu S, Lai S, Wu J, Wu Z, Pang X, Wei X, Jiang X, Zhen X, et al. Fuhrman nuclear grade prediction of clear cell renal cell carcinoma: influence of volume of interest delineation strategies on machine learning-based dynamic enhanced CT radiomics analysis. Eur Radiol. 2022;32(4):2340–50. 10.1007/s00330-021-08322-w . From NLM Medline. 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(1):953. 10.1186/s12885-023-11454-5 . From NLM Medline. Jiao P, Wang B, Ni X, Lu Y, Yang R, Liu Y, Wang J, Mei H, Liu X, Weng X, et al. CT Urography-Based Radiomics to Predict ISUP Grading of Clear Cell Renal Cell Carcinoma. J Cancer. 2025;16(4):1118–26. 10.7150/jca.105173 . From NLM PubMed-not-MEDLINE. Yang Z, Jiang H, Shan S, Wang X, Kou Q, Wang C, Jin P, Xu Y, Liu X, Zhang Y, et al. 2.5D Deep Learning-Based Prediction of Pathological Grading of Clear Cell Renal Cell Carcinoma Using Contrast-Enhanced CT: A Multicenter Study. Acad Radiol. 2025;32(10):5907–16. 10.1016/j.acra.2025.06.056 . From NLM Medline. Huang JL, Liu Q, Wang CL, Lang X, Zeng YX, Zhou DQ. CT-based subregional and peritumoral radiomics for predicting pathological T stage of clear cell renal cell carcinoma: an exploratory study of biological mechanisms. Insights Imaging. 2026;17(1):50. 10.1186/s13244-026-02226-3 . From NLM PubMed-not-MEDLINE. Yang G, Nie P, Yan L, Zhang M, Wang Y, Zhao L, Li M, Xie F, Xie H, Li X, et al. The radiomics-based tumor heterogeneity adds incremental value to the existing prognostic models for predicting outcome in localized clear cell renal cell carcinoma: a multicenter study. Eur J Nucl Med Mol Imaging. 2022;49(8):2949–59. 10.1007/s00259-022-05773-1 . From NLM Medline. Chen S, Gao F, Guo T, Jiang L, Zhang N, Wang X, Zheng J. Deep learning-based multi-model prediction for disease-free survival status of patients with clear cell renal cell carcinoma after surgery: a multicenter cohort study. Int J Surg. 2024;110(5):2970–7. 10.1097/JS9.0000000000001222 From NLM Medline. Song S, Ge W, Qi X, Che X, Wang Q, Wu G. Computed tomography-based radiomics predicts prognostic and treatment-related levels of immune infiltration in the immune microenvironment of clear cell renal cell carcinoma. BMC Med Imaging. 2025;25(1):213. 10.1186/s12880-025-01749-3 . From NLM Medline. Xiong Y, Yao L, Lin J, Yao J, Bai Q, Huang Y, Zhang X, Huang R, Wang R, Wang K, et al. Artificial intelligence links CT images to pathologic features and survival outcomes of renal masses. Nat Commun. 2025;16(1):1425. 10.1038/s41467-025-56784-z . From NLM Medline. Schutte U, Bisht S, Heukamp LC, Kebschull M, Florin A, Haarmann J, Hoffmann P, Bendas G, Buettner R, Brossart P, et al. Hippo signaling mediates proliferation, invasiveness, and metastatic potential of clear cell renal cell carcinoma. Transl Oncol. 2014;7(2):309–21. 10.1016/j.tranon.2014.02.005 . From NLM PubMed-not-MEDLINE. Yuan Z, Li Y, Zhang S, Wang X, Dou H, Yu X, Zhang Z, Yang S, Xiao M. Extracellular matrix remodeling in tumor progression and immune escape: from mechanisms to treatments. Mol Cancer. 2023;22(1):48. 10.1186/s12943-023-01744-8 . From NLM Medline. Guo Y, Gong B, Li Y, Lou J, Shan S, Zhang X, Lu Q, Zhu D, Sun Q, Miao L, et al. Radiomics-Based Unsupervised Clustering Identifies Subtypes Associated With Prognosis and Immune Microenvironment in Clear Cell Renal Cell Carcinoma: A Multicenter Study. Adv Sci (Weinh). 2025;12(34):e06165. 10.1002/advs.202506165 . From NLM Medline. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 19 Apr, 2026 Editor invited by journal 11 Apr, 2026 Editor assigned by journal 10 Apr, 2026 Submission checks completed at journal 10 Apr, 2026 First submitted to journal 02 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9303146","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627013189,"identity":"6bf024fa-c4fe-4a08-a2f8-d1ce6ce37675","order_by":0,"name":"Zhe Jin","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Zhe","middleName":"","lastName":"Jin","suffix":""},{"id":627013190,"identity":"53bc0cdc-9171-47da-9c1b-fc43f0009770","order_by":1,"name":"Siyi Chen","email":"","orcid":"","institution":"Xiangyang No. 1 People's Hospital, Hubei University of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Siyi","middleName":"","lastName":"Chen","suffix":""},{"id":627013193,"identity":"dd95a653-e964-4570-b627-725b4fe8ab5d","order_by":2,"name":"Chen Jin","email":"","orcid":"","institution":"Zhejiang University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Chen","middleName":"","lastName":"Jin","suffix":""},{"id":627013194,"identity":"9e7d4a5e-d92c-4101-b954-3668199798a8","order_by":3,"name":"Ying Ma","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Ma","suffix":""},{"id":627013195,"identity":"d75ddf0a-894d-4b7c-8347-1b8021917360","order_by":4,"name":"Yunshi Liang","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Yunshi","middleName":"","lastName":"Liang","suffix":""},{"id":627013196,"identity":"401a94e4-b527-462c-b38d-d283e52cc231","order_by":5,"name":"Jingxuan Guo","email":"","orcid":"","institution":"Tianjin Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jingxuan","middleName":"","lastName":"Guo","suffix":""},{"id":627013197,"identity":"42ba0f30-7cc2-4e3b-bec5-2b985233c4bc","order_by":6,"name":"Luyi Chen","email":"","orcid":"","institution":"Guangzhou medical university","correspondingAuthor":false,"prefix":"","firstName":"Luyi","middleName":"","lastName":"Chen","suffix":""},{"id":627013198,"identity":"3f222b3c-da08-4d17-a821-7a024e267209","order_by":7,"name":"Yingwen Liu","email":"","orcid":"","institution":"Guangzhou medical university","correspondingAuthor":false,"prefix":"","firstName":"Yingwen","middleName":"","lastName":"Liu","suffix":""},{"id":627013199,"identity":"5a96027b-31cc-4fcf-a234-bf6f827c1437","order_by":8,"name":"Xusheng Lin","email":"","orcid":"","institution":"Guangzhou medical university","correspondingAuthor":false,"prefix":"","firstName":"Xusheng","middleName":"","lastName":"Lin","suffix":""},{"id":627013200,"identity":"f1d4f588-44a8-4f5b-9203-09a58e7c2532","order_by":9,"name":"Chuyi Huang","email":"","orcid":"","institution":"Guangzhou medical university","correspondingAuthor":false,"prefix":"","firstName":"Chuyi","middleName":"","lastName":"Huang","suffix":""},{"id":627013201,"identity":"19f21031-dcdb-4394-b4d9-21274f53c530","order_by":10,"name":"Zhidan Zhong","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Zhidan","middleName":"","lastName":"Zhong","suffix":""},{"id":627013202,"identity":"75c97e5e-a27e-4f79-bceb-6ca96cfdd0f5","order_by":11,"name":"Yongxin Chen","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Yongxin","middleName":"","lastName":"Chen","suffix":""},{"id":627013203,"identity":"43f82905-8c89-49f1-9a5f-03bd4a35ba2e","order_by":12,"name":"Yuan Guo","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Yuan","middleName":"","lastName":"Guo","suffix":""},{"id":627013204,"identity":"67c8532b-08ac-42d3-ac92-1e064d6a1749","order_by":13,"name":"Wenjie Tang","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Wenjie","middleName":"","lastName":"Tang","suffix":""},{"id":627013205,"identity":"67527f6a-f0be-430e-8ae0-a27f1e8dddde","order_by":14,"name":"Weifeng Liu","email":"","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Weifeng","middleName":"","lastName":"Liu","suffix":""},{"id":627013206,"identity":"a07c7f1c-1290-46fa-9e72-f23585db7b3f","order_by":15,"name":"Li Tian","email":"","orcid":"","institution":"State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Tian","suffix":""},{"id":627013207,"identity":"740ac333-d07e-47b3-86a7-a065510daad4","order_by":16,"name":"Xinqing Jiang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8klEQVRIiWNgGAWjYHCChANgir0Byj9AtBYeEJVAnBYokEggUgu/RMLDAz931Cb23Xz+8DPvDwY5vhsJjJ8L8GiRnJGQcLD3zHFjyds5xtI8CQzGkjcSmKVn4NFicCMh4QBv2zE5g9s5bMxALYkbbiQAGXi02AO1HPzbdozH4ObxZyAt9QS1GEgkJBzmbauRM7jBYAbSkmBASIvEmQcJh2XbDhhLnskxlpyTJmE488zDZml8Wvjbc5I/vm2rS+w7fvzhhzc2NvJ8x5MPfsanhUEgJwFIHoZFhwQQMzbg0wC05jhIbR0pkT4KRsEoGAUjDQAAt2pTFX4vO0kAAAAASUVORK5CYII=","orcid":"","institution":"The Second Affiliated Hospital of South China University of Technology (Guangzhou First People’s Hospital)","correspondingAuthor":true,"prefix":"","firstName":"Xinqing","middleName":"","lastName":"Jiang","suffix":""}],"badges":[],"createdAt":"2026-04-02 12:24:57","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9303146/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9303146/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108007593,"identity":"13c7b633-a8d6-4c5e-a4ef-8eeb7488837d","added_by":"auto","created_at":"2026-04-28 13:00:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":85072,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFlowchart of patient recruitment. \u003c/strong\u003eThis study included four cohorts: the SYUCC cohort (training and internal validation) and three external validation cohorts, namely the GZFPH cohort, the KiTS19 cohort, and the TCGA cohort.\u003c/p\u003e\n\u003cp\u003eNote: ccRCC = clear cell renal cell carcinoma.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/3d9fc85b7d885f3934780cbf.png"},{"id":107948552,"identity":"19f6adb9-7396-4ac5-b413-88b7cc4e7883","added_by":"auto","created_at":"2026-04-28 00:22:02","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":225809,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe model analysis workflow. \u003c/strong\u003eThe study consists of four key phases: (1) Image preprocessing. (2) Tumor habitat construction based on supervoxel segmentation and prototype alignment. (3) Extraction of habitat heterogeneity and conventional radiomics features. (4) Feature selection. (5) Development of Habitat, Radiomics, and Combined models for WHO/ISUP grade prediction.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/008901c9daf23187f982e581.png"},{"id":107948557,"identity":"2ca98b94-6f91-4d4c-81f9-8eb38474525f","added_by":"auto","created_at":"2026-04-28 00:22:02","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":61483,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredictive performance of the Radiomics, Habitat, and Combined models for WHO/ISUP grading across cohorts. \u003c/strong\u003e(a-d) ROC curves of the three models in the SYUCC cohort (a), GZFPH cohort (b), KiTS19 cohort (c), and TCGA-KIRC cohort (d), respectively.\u003c/p\u003e\n\u003cp\u003eNote: ROC = receiver operating characteristic; AUC = area under the ROC curve; ccRCC = clear cell renal cell carcinoma.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/d1cd0539f0f7e94aa5b156b5.png"},{"id":108006557,"identity":"03e47845-8d74-48c9-a635-a2756f856784","added_by":"auto","created_at":"2026-04-28 12:56:00","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":78682,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrognostic value of WHO/ISUP grade and model-derived risk stratification.\u003c/strong\u003e Kaplan-Meier overall survival (OS) curves were generated for the KiTS19 and TCGA cohorts using (a, c) the ground-truth WHO/ISUP grade (true low-grade vs true high-grade) and (b, d) the model-derived risk stratification (low-risk vs high-risk), respectively. Panels show results for the KiTS19 cohort based on ground truth (a) and model prediction (b), and for the TCGA cohort based on ground truth (c) and model prediction (d). P values were calculated using the log-rank test.\u003c/p\u003e\n\u003cp\u003eNote: OS = overall survival; WHO/ISUP = World Health Organization/International Society of Urological Pathology.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/3081cfa89465beffc1d9ac26.png"},{"id":108007020,"identity":"466486ad-fa7a-4722-a00d-07c8e53e8aec","added_by":"auto","created_at":"2026-04-28 12:58:15","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":280496,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDifferential gene expression and pathway enrichment associated with risk groups derived from the Combined model. \u003c/strong\u003e(a) Volcano plot of differentially expressed genes between the high-risk and low-risk groups. (b) Bar plot of representative genes upregulated in the high-risk group, ranked by log2 fold change. (c, d) GSEA using the C2 (curated gene sets) database showing significantly upregulated (c) and downregulated (d) pathways between WHO/ISUP grades as predicted by the combined model. (e) Bubble plot depicting correlations between habitat features and tumor biological processes.\u003c/p\u003e\n\u003cp\u003eNote: GSEA = Gene Set Enrichment Analysis; NES = normalized enrichment score.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/0c5b5fda1215ed9c074545bb.jpeg"},{"id":108006453,"identity":"d8a47f89-3b06-42d6-9e42-fa5d4df58144","added_by":"auto","created_at":"2026-04-28 12:55:38","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":55959,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImmune microenvironment characteristics between model-derived risk groups. \u003c/strong\u003e(a) Comparison of IPS-related immune components between low- and high-risk groups, including MHC, EC, SC, and CP scores. (b) Differences in tumor-infiltrating immune cell abundance between groups estimated using MCPcounter test.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/2f06fbea915ca3e59caca5bf.png"},{"id":107948555,"identity":"11a220ac-49a2-43d3-9ad1-33e6d9c7b94c","added_by":"auto","created_at":"2026-04-28 00:22:02","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":347210,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRepresentative habitat visualizations for WHO/ISUP grading prediction. \u003c/strong\u003e(A) A ccRCC patient with high WHO/ISUP grade. (B) A ccRCC patient with low WHO/ISUP grade. For each case, axial and coronal contrast-enhanced CT images with magnified tumor ROI are shown with the corresponding habitat maps and 3D habitat reconstructions. The model-predicted WHO/ISUP grade is displayed on the right. Habitat subregions are color-coded as S1 (red), S2 (green), and S3 (blue).\u003c/p\u003e\n\u003cp\u003eNote: WHO/ISUP = World Health Organization/International Society of Urological Pathology; ccRCC = clear cell renal cell carcinoma; ROI = region of interest.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/3d6efb6336fdde753558b8e6.png"},{"id":108008934,"identity":"d42d078d-567e-47a4-8930-9a753b764876","added_by":"auto","created_at":"2026-04-28 13:08:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1566208,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/ec45e8c4-8375-4944-b541-b4274d06012c.pdf"},{"id":107948551,"identity":"c692c343-880d-439c-ac66-a60097c75509","added_by":"auto","created_at":"2026-04-28 00:22:02","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":16186,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-9303146/v1/6727b91081f907d19d84baf6.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Preoperative CT Habitat Analysis for Predicting WHO/ISUP Grade in Clear Cell Renal Cell Carcinoma: A Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eClear cell renal cell carcinoma (ccRCC) is the most common histological subtype of renal cell carcinoma, accounting for approximately 70% of cases and contributing substantially to kidney cancer\u0026ndash;related mortality\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. The WHO / International Society of Urological Pathology (ISUP) grading system is a critical prognostic determinant in ccRCC, closely associated with tumor aggressiveness, risk of recurrence and metastasis, and long-term survival\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Compared with low-grade tumors (grades I\u0026ndash;II), high-grade tumors (grades III\u0026ndash;IV) generally indicate more aggressive tumor biology and poorer clinical outcomes\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Accordingly, accurate assessment of WHO / ISUP grade is essential for individualized clinical management, including surgical planning, perioperative risk stratification, and the tailoring of postoperative surveillance and adjuvant treatment strategies\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. However, reliable preoperative determination of WHO/ISUP grade remains challenging. Percutaneous renal tumour biopsy is a commonly used method for preoperative grading of ccRCC, yet it is invasive and prone to sampling error, particularly in the presence of marked intratumoral heterogeneity, which can lead to grade underestimation or discordant results\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. These limitations constrain its preoperative utility and underscore the necessity of a noninvasive, reproducible imaging approach for WHO/ISUP grade prediction to support risk-adapted decision-making in ccRCC.Computed tomography (CT) is the most widely used imaging modality for preoperative evaluation of renal masses, playing a central role in tumor localization, characterization of enhancement patterns, and preoperative staging\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. With the advancement of radiomics, quantitative image analysis enables the extraction of high-dimensional features that capture tumor morphology, intensity distribution, and textural heterogeneity, thereby enabling noninvasive preoperative prediction of pathological characteristics\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. In recent years, several studies have developed CT- or MRI-based radiomics models to predict the nuclear grade of ccRCC\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. However, most existing models rely on whole-tumor regions of interest (ROIs) for feature extraction. Such global characterization may obscure regional differences by averaging heterogeneous tumor components, thereby limiting the ability to fully capture intratumoral spatial heterogeneity (ITH)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. This spatial heterogeneity may reflect distinct molecular alterations, proliferative activity, and malignant potential across tumor subregions\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Moreover, the biological underpinnings of radiomic features remain insufficiently elucidated, which in turn restricts model interpretability and limits robust generalization across institutions\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo better characterize intratumoral spatial heterogeneity, recent studies in several solid tumors, including breast cancer, have increasingly explored unsupervised subregional partitioning strategies\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. This strategy automatically partitions tumors into subregions with similar imaging characteristics based on voxel-level intensity and texture distributions, and subsequently quantifies spatial organization using ecological diversity or heterogeneity indices. Habitat-based modeling therefore enables a more refined characterization of ITH and has demonstrated potential to enhance predictive performance and interpretability. ccRCC typically exhibits marked spatial heterogeneity on imaging, with substantial variations in enhancement patterns, grayscale distribution, and textural architecture across tumor regions\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Such heterogeneity may reflect the complex \u0026ldquo;ecological architecture\u0026rdquo; of the tumor and correspond to distinct biological states and malignant potential. However, the role of ITH-driven subregional modeling in predicting WHO/ISUP grade in ccRCC remains insufficiently investigated.\u003c/p\u003e \u003cp\u003eIn addition, the clinical translation of radiomics models fundamentally depends on their biological interpretability\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Emerging evidence from radiogenomic studies suggests that imaging phenotypes may partially reflect underlying molecular characteristics, metabolic reprogramming, angiogenesis, and the immune microenvironment, thereby enabling the establishment of verifiable links between imaging features and tumor biology\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. However, in the context of ccRCC nuclear grade prediction, most prior studies have focused on constructing discriminative models and reporting performance metrics, with limited investigation into the biological pathways and immunological correlates underlying key radiomic features\u0026mdash;particularly those related to intratumoral spatial heterogeneity. Leveraging external cohorts with transcriptomic data to conduct pathway enrichment and immune infiltration analyses is therefore crucial for elucidating the biological basis of ITH-associated imaging phenotypes and strengthening model credibility and translational applicability.\u003c/p\u003e \u003cp\u003eTherefore, in this study, we aimed to evaluate the performance and clinical utility of CT-derived intratumoral spatial heterogeneity for preoperative prediction of WHO/ISUP grade and prognostic risk stratification in ccRCC. Furthermore, by leveraging an external cohort with available transcriptomic data, we performed radiogenomic and immune-related analyses to investigate the molecular pathways and immune microenvironment characteristics associated with ITH imaging phenotypes. This integrative approach may provide biological insights into the proposed model and enhance its clinical credibility and translational potential.\u003c/p\u003e"},{"header":"Method and Materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ePatients and study design\u003c/h2\u003e \u003cp\u003e This multicenter retrospective study was approved by the institutional review boards of all participating hospitals and was conducted in accordance with the Declaration of Helsinki. Due to its retrospective nature, written informed consent for retrospective datasets was waived. Four cohorts were included, comprising two institutional cohorts and two public datasets. The institutional cohorts were obtained from Sun Yat-sen University Cancer Center (SYUCC) and Guangzhou First People\u0026rsquo;s Hospital (GZFPH). The inclusion criteria were: (1) pathologically confirmed ccRCC after surgery; (2) no concurrent other malignant tumors; (3) underwent radical or partial nephrectomy; and (4) available preoperative abdominal CT. The exclusion criteria were: (1) absence of pathological WHO/ISUP nuclear grade information; (2) lack of contrast-enhanced CT or poor image quality; and (3) incomplete clinical data. In the institutional cohorts, 308 patients from SYUCC and 106 patients from GZFPH were finally included.\u003c/p\u003e \u003cp\u003eIn addition, two public datasets were included: KiTS19 (Kidney Tumor Segmentation 2019 [KiTS19]; MICCAI 2019 Kidney Tumor Segmentation Challenge; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/neheller/kits19\u003c/span\u003e\u003cspan address=\"https://github.com/neheller/kits19\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) and TCGA-KIRC (The Cancer Genome Atlas [TCGA]; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.cancer.gov/tcga\u003c/span\u003e\u003cspan address=\"https://www.cancer.gov/tcga\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e; TCGA-KIRC). For KiTS19, 210 patients were initially identified; after excluding 67 patients with non-ccRCC histology and 1 patient with poor image quality, 142 patients were included. For TCGA-KIRC, patients without contrast-enhanced CT, with poor image quality, or with incomplete clinical information were excluded, resulting in 148 ccRCC patients. The detailed recruitment flowcharts are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn total, 704 patients were included. The SYUCC cohort was randomly divided into training and internal validation sets at a 8:2 ratio for model development. The GZFPH, KiTS19, and TCGA-KIRC cohorts served as independent external validation cohorts.\u003c/p\u003e \u003cp\u003eThe study was conducted in four phases: Phase 1: Model development; Phase 2: Multi-center validation; Phase 3: Prognostic prediction analysis; Phase 4: Biological interpretability analysis.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePathological assessment and clinical outcomes\u003c/h3\u003e\n\u003cp\u003ePathological data were retrieved from postoperative pathology reports in the hospital information system. Pathological diagnosis and grading were assessed by experienced senior pathologists according to the 2016 WHO classification criteria\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. For analyses, WHO/ISUP grades I\u0026ndash;II were categorized as low grade, and grades III\u0026ndash;IV as high grade. Clinical and pathological variables collected included age, sex, tumor laterality, tumor size, clinical T stage, and ISUP grade.\u003c/p\u003e \u003cp\u003eFollow-up information for the KiTS19 and TCGA-KIRC cohorts has been previously reported in the corresponding public dataset documentation and related studies\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Overall survival (OS) data were extracted to evaluate the prognostic stratification performance of the proposed model. All 142 patients included from KiTS19 had available OS data, and 128 patients in the TCGA-KIRC cohort had available OS data. OS was defined as the time from initial surgery to death from any cause, or to the last follow-up for censored patients.\u003c/p\u003e\n\u003ch3\u003eCT imaging acquisition\u003c/h3\u003e\n\u003cp\u003eAll patients underwent preoperative contrast-enhanced abdominal CT examinations within 2 weeks prior to surgery. CT images were acquired at multiple institutions using multi-detector CT scanners, including Siemens, Toshiba, GE, and Philips systems.\u003c/p\u003e \u003cp\u003eFour-phase CT scans were routinely performed, including unenhanced phase, corticomedullary phase, nephrographic phase, and excretory phase. Nonionic iodinated contrast medium was intravenously administered using a power injector at a dose of 1\u0026ndash;1.5 mL/kg (maximum 150 mL), with an injection rate of 2.5\u0026ndash;3.5 mL/s.\u003c/p\u003e \u003cp\u003eFor model development, the corticomedullary phase images were selected for subsequent analysis, based on previous evidence suggesting that this phase provides optimal tumor\u0026ndash;renal parenchyma contrast and better differentiation between high-grade and low-grade ccRCC\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDetailed CT acquisition parameters for each institution are summarized in \u003cb\u003eSupplementary Material 1\u003c/b\u003e.\u003c/p\u003e\n\u003ch3\u003eImage preprocessing and subregion segmentation\u003c/h3\u003e\n\u003cp\u003eThe overall workflow is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, and detailed procedures are described below. Tumor subregions were generated using a two-stage unsupervised strategy. Two experienced radiologists (T.W.J. and G.Y., with 10 and 18 years of experience in abdominal imaging, respectively) manually delineated the three-dimensional tumor volume of interest (VOI) on contrast-enhanced CT using ITK-SNAP (v3.8.0). To enhance cross-center consistency, CT images were first resampled to isotropic voxel spacing (1\u0026times;1\u0026times;1 mm\u0026sup3;). Within each tumor VOI, voxel intensities were clipped using percentile-based truncation to mitigate scanner-dependent extreme values. Robust intensity normalization was subsequently applied to reduce the influence of residual outliers, followed by z-score standardization to harmonize feature scales across patients.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn Stage 1, mask-constrained three-dimensional supervoxels were generated using the SLIC algorithm implemented in SimpleITK. Supervoxel size was set to 8\u0026times;8\u0026times;8 voxels in isotropic space, and the compactness parameter was fixed at 0.4 to balance spatial regularity and boundary adherence, thereby producing spatially contiguous and anatomically coherent initial units. In Stage 2, a predefined 20-dimensional feature vector was extracted from each supervoxel, primarily consisting of first-order intensity statistics. During model training, these feature vectors were standardized and mapped into a global prototype space constructed exclusively from training-set samples. Supervoxels were assigned to prototype categories via nearest-neighbor matching. During inference, weighted K-means clustering (weighted by supervoxel volume) was performed within each case to obtain N subregions. The optimal number of subregions was determined using bayesian information criterion analysis on the training set, which indicated N\u0026thinsp;=\u0026thinsp;3 as the most parsimonious and stable solution. Case-specific clusters were then aligned with the global prototype labels using a one-to-one greedy matching strategy to ensure consistent, non-overlapping subregion labeling across patients. This procedure yielded refined tumor habitats characterized by improved internal homogeneity and cross-patient label consistency.\u003c/p\u003e\n\u003ch3\u003eModel development\u003c/h3\u003e\n\u003cp\u003eWe first applied minimum redundancy maximum relevance to pre-screen the initial imaging features and reduce redundancy in the high-dimensional feature space. For both intratumoral habitat heterogeneity features and conventional radiomics features, feature repeatability was evaluated prior to subsequent selection. Specifically, five pseudo-masks were generated to simulate contour perturbations of manual delineation, and features were repeatedly extracted from each pseudo-mask to calculate intraclass correlation coefficients (ICCs). Features were considered repeatable and retained for downstream analyses if ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.90 for habitat features and ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.75 for radiomics features (a stricter ICC threshold was used for habitat features because they rely more on subregion partitioning and are more sensitive to boundary perturbations). Next, low-variance filtering was performed to remove uninformative features, followed by Pearson correlation analysis to eliminate highly correlated features (|r| \u0026gt; 0.75) and mitigate multicollinearity. Candidate features significantly associated with the endpoint were then identified using an independent-samples t test or the Mann\u0026ndash;Whitney U test, as appropriate (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003eBased on the above statistical filtering steps, a second-stage feature selection was conducted using a genetic algorithm combined with a support vector machine (SVM) to obtain a more robust and generalizable feature subset. The training set was split into a sub-training set and an internal validation set. Each GA individual represented a fixed-length feature subset, and the fitness function was defined by the classification performance of an SVM model (AUC and accuracy) on both the sub-training and internal validation sets. The optimal feature subset was selected and saved. Finally, 20 habitat features and 20 radiomics features were determined for model construction (\u003cb\u003eSupplementary Material 2\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eUsing the final feature sets, three predictive models were developed: (1) a Radiomics model, (2) a Habitat model, and (3) a Combined model. Both the Radiomics and Habitat models were built using logistic regression (LR) to generate their respective rad-scores. For the Combined model, score-level fusion was implemented by feeding the two rad-scores into a second-stage logistic regression model to derive a combined rad-score for final classification. All model development was implemented in Python (scikit-learn v1.0.2), and model selection and performance evaluation were conducted using the training and internal validation sets.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel assessment\u003c/h2\u003e \u003cp\u003eAll models were developed and evaluated using Python (v3.7.7) with scikit-learn (v1.0.2). Hyperparameter tuning was performed exclusively within the training set using grid search combined with five-fold cross-validation, and the model with the best mean cross-validated performance was selected. The finalized models were subsequently evaluated on the internal validation set and three independent external cohorts (GZFPH, KiTS19, and TCGA-KIRC).\u003c/p\u003e \u003cp\u003eTo assess the incremental value of different feature sets, we constructed and compared a Radiomics model based on conventional radiomics features, a Habitat model derived from intratumoral subregion (habitat) characteristics, and a Combined model integrating both feature types. Discrimination for WHO/ISUP grade (high vs low) was evaluated using receiver operating characteristic (ROC) curves and quantified by the area under the ROC curve (AUC). Sensitivity, specificity and accuracy were c alculated at a prespecified threshold. AUCs were compared using the DeLong test, and 95% confidence intervals were estimated via bootstrap resampling.\u003c/p\u003e \u003cp\u003eTo enhance model interpretability, intratumoral habitat segmentation maps were spatially visualized. Subregion labels were projected back onto the original CT images and overlaid using color-coded representations to illustrate spatial heterogeneity patterns within tumors. Representative cases with true-positive, true-negative, false-positive, and false-negative predictions were presented to qualitatively demonstrate the relationship between spatial subregion distribution and model predictions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eBiological interpretability analysis\u003c/h3\u003e\n\u003cp\u003eTo explore the biological underpinnings associated with the model-derived risk stratification, we performed a multi-level biological interpretability analysis at the transcriptomic level using TCGA ccRCC cohort.\u003c/p\u003e \u003cp\u003eGene expression data and corresponding clinical information were obtained from the TCGA database. Patients were stratified into high- and low-grade groups according to the model-derived risk score. Differential pathway enrichment analysis between groups was conducted using Gene Set Enrichment Analysis (GSEA). The analysis was implemented with the clusterProfiler, msigdbr, and GSEAbase R packages based on the c2.cp and Reactome gene sets from the Molecular Signatures Database. Enriched pathways were identified using the criteria of nominal P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 and absolute normalized enrichment score (|NES|)\u0026thinsp;\u0026gt;\u0026thinsp;1. To further quantify pathway-level activity in individual samples, Gene Set Variation Analysis (GSVA) was performed using the GSVA R package. The limma package was subsequently applied to assess differential pathway activity between groups. Gene sets were considered significantly different with P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 and |log2 fold change| \u0026gt; 0.1. To evaluate the association between deep learning-derived imaging features and biological pathways, Spearman correlation analysis was conducted between model features and GSVA-derived pathway scores. Correlations with P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were considered statistically significant.\u003c/p\u003e \u003cp\u003eTumor microenvironment (TME) cellular composition was estimated using single-sample Gene Set Enrichment Analysis (ssGSEA) implemented in the GSVA package. Cell-type-specific enrichment scores were calculated for each sample, and differences in immune cell abundance between risk groups were assessed using nonparametric statistical tests.\u003c/p\u003e\n\u003ch3\u003eSurvival analysis\u003c/h3\u003e\n\u003cp\u003eSurvival analyses were conducted in the KiTS19 and TCGA cohorts. OS was estimated using the Kaplan\u0026ndash;Meier method and compared using the log-rank test. Within each cohort, OS was analyzed under two stratification schemes: (i) the ground-truth WHO/ISUP grade (true low-grade vs true high-grade) and (ii) the model-derived risk groups. For the model-based stratification, patients were dichotomized into low- and high-risk groups according to the cohort-specific median value of the Combined model\u0026ndash;derived rad-score, and Kaplan-Meier curves were generated accordingly.\u003c/p\u003e \u003cp\u003eAll statistical tests were two-sided, and P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant. Statistical analyses were performed using R (version 3.4.1), SPSS (version 23.0; IBM, Armonk, NY, USA), and Python (version 3.12).\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eClinicopathologica characteristics\u003c/h2\u003e \u003cp\u003eA total of 704 patients with pathologically confirmed ccRCC were included, comprising 483 low-grade (WHO/ISUP I\u0026ndash;II) and 221 high-grade (III\u0026ndash;IV) tumors. The SYUCC cohort (n\u0026thinsp;=\u0026thinsp;308) served as the development set, and the GZFPH (n\u0026thinsp;=\u0026thinsp;106), KiTS19 (n\u0026thinsp;=\u0026thinsp;142), and TCGA-KIRC (n\u0026thinsp;=\u0026thinsp;148) cohorts were used for external validation. Baseline clinicopathologic characteristics across the four cohorts are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline clinicopathologic characteristics of patients with ccRCC in the four cohorts.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSYUCC cohort (N\u0026thinsp;=\u0026thinsp;308)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGZFPH cohort (N\u0026thinsp;=\u0026thinsp;106)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKiTS19 cohort (N\u0026thinsp;=\u0026thinsp;142)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTCGA cohort (N\u0026thinsp;=\u0026thinsp;148)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e52.0\u0026thinsp;\u0026plusmn;\u0026thinsp;12.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59\u0026thinsp;\u0026plusmn;\u0026thinsp;12.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e59.3\u0026thinsp;\u0026plusmn;\u0026thinsp;12.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e60.4\u0026thinsp;\u0026plusmn;\u0026thinsp;12.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e219(71.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71(67.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e92(64.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e89(60.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e89(28.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35(33.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e50(35.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e59(39.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSide, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e147(47.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49(46.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e70(49.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e79(53.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e161(52.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e57(53.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e72(50.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e69(46.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaximum tumour diameter, cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.8\u0026thinsp;\u0026plusmn;\u0026thinsp;3.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.2\u0026thinsp;\u0026plusmn;\u0026thinsp;3.0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT staging, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e267(86.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81(76.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e100(70.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e87(58.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20(6.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10(9.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7(4.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14(9.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21(6.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13(12.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e34(23.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e45(30.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0(0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2(1.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1(0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2(1.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWHO/ISUP grade, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25(8.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23(21.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e19(13.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1(0.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e205(66.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e65(61.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e82(57.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63(42.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e68(22.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12(11.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e31(21.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e61(41.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6(5.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10(7.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e23(15.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eData were presented as number of patients, with the exception of age and maximum tumour diamete at baseline (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD).\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eAbbreviations: ccRCC\u0026thinsp;=\u0026thinsp;clear cell renal cell carcinoma; ISUP\u0026thinsp;=\u0026thinsp;International Society of Urological Pathology; SD= standard deviation.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAcross cohorts, the mean age ranged from 52.0 to 60.4 years. Males represented the majority of patients in each cohort, and tumor laterality was generally balanced between the right and left kidneys. The mean maximum tumor diameter varied across cohorts, ranging from 4.2 cm in the SYUCC cohort to 6.2 cm in the TCGA cohort. Most tumors were stage T1 in all cohorts, although the proportion of T1 disease was lower in the TCGA and KiTS19 cohorts than in the SYUCC and GZFPH cohorts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePredictive performance of the models for WHO/ISUP grading\u003c/h2\u003e \u003cp\u003eThe predictive performance of the Radiomics, Habitat, and Combined models for WHO/ISUP grade classification is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea\u0026ndash;d. Overall, the Combined model showed the highest performance across cohorts, with AUCs of 0.860 (95% CI: 0.813\u0026ndash;0.907) in the SYUCC cohort, 0.830 (0.813\u0026ndash;0.907) in the GZFPH cohort, 0.829 (0.756\u0026ndash;0.901) in the KiTS19 cohort, and 0.750 (0.664\u0026ndash;0.830) in the TCGA cohort. Although the AUC differences between the Combined and Habitat models were not statistically significant across cohorts (P\u0026thinsp;\u0026gt;\u0026thinsp;0.05), the Combined model demonstrated higher overall accuracy in all datasets (0.805, 0.764, 0.796, and 0.730, respectively), together with favorable sensitivity and specificity trade-offs in several cohorts. The Habitat model also demonstrated strong discriminative performance, with consistently higher AUCs than the Radiomics model across all cohorts (SYUCC: 0.830 vs 0.800; GZFPH: 0.814 vs 0.797; KiTS19: 0.812 vs 0.792; TCGA: 0.750 vs 0.610). Notably, in the TCGA cohort, both the Habitat and Combined models improved discrimination compared with the Radiomics model (AUC 0.750 vs 0.610).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003ePerformance of radiomics, habitat, and combined models for WHO/ISUP grade prediction across four cohorts.\u003c/b\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCohorts\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSYUCC cohort\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.800 (0.740\u0026ndash;0.860)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.779\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.770\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.783\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.013\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHabitat\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.830 (0.775\u0026ndash;0.885)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.072\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.860 (0.813\u0026ndash;0.907)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.816\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRef\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eGZFPH cohort\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.797 (0.707\u0026ndash;0.877)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.736\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.737\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.736\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.146\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHabitat\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.814 (0.735\u0026ndash;0.888)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.708\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.947\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.655\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.575\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.830 (0.813\u0026ndash;0.907)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.764\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.842\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRef\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eKiTS19 cohort\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.792 (0.707\u0026ndash;0.877)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.732\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.756\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.723\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.060\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHabitat\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.812 (0.735\u0026ndash;0.888)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.725\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.878\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.663\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.593\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.829 (0.756\u0026ndash;0.901)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.780\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRef\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eTCGA cohort\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.610 (0.519\u0026ndash;0.701)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.608\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.595\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.006\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHabitat\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.750 (0.668\u0026ndash;0.834)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.716\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.719\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.496\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.750 (0.664\u0026ndash;0.830)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.730\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.774\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.672\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRef\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003eAbbreviations: AUC\u0026thinsp;=\u0026thinsp;area under the receiver operating characteristics curve; 95% CI\u0026thinsp;=\u0026thinsp;95% confidence interval.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eSurvival prediction\u003c/h2\u003e \u003cp\u003eKaplan-Meier analyses were performed in the KiTS19 and TCGA cohorts to evaluate OS stratification by (i) the ground-truth WHO/ISUP grade and (ii) the model-derived risk groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). In the KiTS19 cohort, OS differed significantly between true low-grade and true high-grade tumors (log-rank p\u0026thinsp;=\u0026thinsp;0.012), and a significant separation was also observed between the model-defined low-risk and high-risk groups (log-rank p\u0026thinsp;=\u0026thinsp;0.017). In the TCGA cohort, OS stratification by the ground-truth WHO/ISUP grade did not reach statistical significance (log-rank p\u0026thinsp;=\u0026thinsp;0.065), whereas the model-derived risk stratification showed significant separation between the low-risk and high-risk groups (log-rank p\u0026thinsp;=\u0026thinsp;0.0032), with the high-risk group exhibiting consistently lower survival probability over follow-up.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eGene expression and pathway enrichment analysis\u003c/h2\u003e \u003cp\u003eTo characterize transcriptomic differences associated with the Combined model derived risk stratification, we compared differential expression profiles between the high risk and low risk groups. The volcano plot in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea summarizes the differentially expressed genes (DEGs) between the two groups, and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb highlights representative genes upregulated in the high risk group, including IL6, TREM1, WNT10B, COL7A1, and SCO2 (ranked by log2 fold change). Functionally, IL6 and TREM1 are linked to inflammatory and innate immune signaling, COL7A1 is related to extracellular matrix components, WNT10B is involved in WNT associated transcriptional regulation, and SCO2 is associated with mitochondrial energy metabolism.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAt the gene set level, GSEA based on the MSigDB C2 (curated gene sets) collection showed that gene sets significantly upregulated in the high risk group were primarily enriched for immune effector and proliferation related processes (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec), exemplified by the natural killer cell related gene set (CURSONS_NATURAL_KILLER_CELLS) and a cancer proliferation and grade related gene set (CANCER_PROLIFERATION_CLUSTER). In contrast, significantly downregulated gene sets were mainly related to epithelial differentiation and EMT and TGF-β associated transcriptional programs (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed), including the epithelial differentiation module (BOSCO_EPITHELIAL_DIFFERENTIATION_MODULE) and a TGF-β mediated EMT gene set (FOROUTAN_TGFB_EMT_DN).\u003c/p\u003e \u003cp\u003eFurthermore, we quantified pathway activities using GSVA and assessed correlations between habitat features and pathway scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee; red indicates positive correlations and blue indicates negative correlations). Overall, habitat features were significantly associated with immune effector related pathway modules, represented by antigen presentation and cross presentation and TCR associated signaling, and with cell cycle and DNA replication and proliferation modules, including S phase, G2 and M checkpoints, and DNA replication. At the subregion level, S1 related habitat features showed an overall positive correlation with the immune effector module and the cell cycle and replication module, whereas S3 related features more frequently exhibited correlations in the opposite direction (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eImmune microenvironment\u003c/h2\u003e \u003cp\u003eBuilding on the transcriptomic analyses that revealed immune-related differences between the model-derived risk groups, we further characterized the immune microenvironment using IPS components and MCPcounter-based deconvolution (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). In the IPS analysis, all four components showed significant between-group differences. Specifically, the high-risk group exhibited higher MHC_IPS (an antigen-presentation-related component; p\u0026thinsp;=\u0026thinsp;0.0009) and higher EC_IPS (an effector-cell-related component; p\u0026thinsp;=\u0026thinsp;0.0094) (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). In addition, SC_IPS (a suppressor-cell-related component; p\u0026thinsp;=\u0026thinsp;0.0445) and CP_IPS (a checkpoint-related component; p\u0026thinsp;=\u0026thinsp;0.0002) also differed significantly between the two groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eConsistent with these score-level differences, MCPcounter deconvolution identified significant group-level differences in inferred immune cell infiltration (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb). Compared with the low-risk group, the high-risk group showed higher estimated abundance of cytotoxic lymphocytes (p\u0026thinsp;=\u0026thinsp;0.0077), neutrophils (p\u0026thinsp;=\u0026thinsp;0.0132), NK cells (p\u0026thinsp;=\u0026thinsp;0.0293), and CD8\u0026thinsp;+\u0026thinsp;T cells (p\u0026thinsp;=\u0026thinsp;0.0283). Other MCPcounter signatures did not show significant differences. Collectively, these findings indicate that the model-derived risk stratification is associated with distinct immune-related profiles at both the IPS score level and the cell-composition level.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eModel visualization\u003c/h2\u003e \u003cp\u003eTo provide an intuitive visualization of intratumoral spatial heterogeneity patterns associated with WHO/ISUP grade prediction, we present representative cases with habitat subregion maps overlaid on the tumor ROI and corresponding 3D reconstructions (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Patient A (predicted high grade) and Patient B (predicted low grade) exhibited distinct intratumoral habitat configurations. In Patient A, S1 (red) was visually more prominent across both 2D slices and the 3D reconstruction, whereas Patient B showed a more prominent S3 (blue). Overall, these qualitative examples illustrate that model predictions align with distinct S1/S3 spatial heterogeneity patterns and are in line with the pathway-activity mapping results (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we proposed and validated a preoperative CT-based framework for predicting WHO/ISUP grade in ccRCC. Unlike conventional modeling strategies that treat tumors as homogeneous entities, our approach incorporates intratumoral habitat subregions to characterize spatial heterogeneity. The resulting combined model demonstrated stable discriminative performance across multicenter datasets and two public cohorts, with AUCs of at least 0.750 in all cohorts. At the molecular level, the model-derived risk stratification corresponded to transcriptomic differences mainly involving immune effector activity, cell proliferation, and epithelial differentiation/EMT-related pathways. Further analyses indicated that habitat subregion features were associated with these pathway modules, thereby linking \u0026ldquo;spatial subregional phenotypes\u0026rdquo; with \u0026ldquo;molecular pathway themes\u0026rdquo; and enabling intuitive visualization of subregional tumor architecture. In addition, stratification based on the model-derived risk score effectively separated overall survival in both the KiTS19 and TCGA cohorts, suggesting that this imaging phenotype not only reflects pathological grading information but also captures prognostically relevant cues of biological aggressiveness.\u003c/p\u003e \u003cp\u003eIn this study, we introduced a habitat-based framework for feature representation to explicitly quantify intratumoral spatial heterogeneity, rather than modeling tumors solely with whole-lesion averaged phenotypes. This design allows imaging signatures to better reflect the underlying structural and functional differences within the tumor. Across the multicenter cohorts, the Habitat model generally outperformed the Radiomics model, and its AUC did not differ significantly from that of the Combined model. This suggests that subregional heterogeneity alone provides strong discriminative signals for WHO/ISUP grade prediction. Meanwhile, the Combined model consistently achieved higher accuracy in most cohorts, indicating a decision-level gain: building upon the core heterogeneity information captured by habitat features, integrating complementary representations may further stabilize classification and better align with the clinical need for correct grading. Notably, the models maintained relatively consistent discriminative performance across different datasets, supporting their robustness and potential generalizability across heterogeneous multicenter cohorts.\u003c/p\u003e \u003cp\u003ePrior CT-based radiomics and deep learning studies have reported encouraging performance for predicting ccRCC WHO/ISUP grade\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan additionalcitationids=\"CR25 CR26\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. However, many approaches still rely predominantly on whole-tumor features, which may obscure critical intratumoral heterogeneity\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan additionalcitationids=\"CR25 CR26\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. In addition, some earlier studies were limited by insufficient external validation or small test cohorts, raising concerns regarding generalizability\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. In contrast, subregional strategies have been more frequently explored in ccRCC tasks related to tumor aggressiveness, with evidence suggesting that partitioning tumors into distinct subregions can better characterize spatial heterogeneity and, in some settings, improve performance over whole-tumor analysis\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Our habitat framework follows this line of reasoning and further applies it to WHO/ISUP grading, a clinically risk-relevant pathologic stratification metric, enabling heterogeneous imaging phenotypes to contribute to grade prediction in a reproducible manner and to be validated across independent cohorts.\u003c/p\u003e \u003cp\u003eBeyond grading prediction performance, we further evaluated the association between the model-derived risk stratification and OS in the KiTS19 and TCGA cohorts, and compared it with OS stratification based on the ground-truth WHO/ISUP grade. In the KiTS19 cohort, the ground-truth WHO/ISUP grade significantly separated OS (log-rank p\u0026thinsp;=\u0026thinsp;0.012), supporting its prognostic stratification value in this cohort. In contrast, OS stratification by the ground-truth grade in the TCGA cohort showed only a borderline difference (p\u0026thinsp;=\u0026thinsp;0.065). Notably, despite the variability in prognostic separation by ground-truth grading across cohorts, the model-derived risk groups consistently discriminated OS in both datasets (KiTS19: p\u0026thinsp;=\u0026thinsp;0.017; TCGA: p\u0026thinsp;=\u0026thinsp;0.0032), demonstrating a more uniform prognostic stratification trend\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. These findings suggest that contrast-enhanced CT\u0026ndash;based imaging phenotypes, particularly those capturing spatial heterogeneity, can provide prognostic support in external public cohorts and may offer preoperative survival-related risk cues, thereby serving as a complementary imaging-derived source of information alongside the WHO/ISUP grading system\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo further elucidate the tumor biological implications underlying this imaging-based stratification, we interpreted the model-derived risk groups from a transcriptomic perspective. Overall, the high-risk group exhibited relatively clear thematic axes. First, programs related to immune effector function and inflammation were more active, exemplified by natural killer cell\u0026ndash;related pathways and FOXP3-associated immune regulatory programs, accompanied by enhanced cell cycle and proliferation signals\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. Second, cell-state programs related to epithelial differentiation and EMT/TGF-β showed a coordinated downregulation, including the epithelial differentiation module, ZEB1 target gene sets, and multiple TGF-β\u0026ndash;mediated EMT pathways\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. At the DEG level, IL6 and TREM1 point to an inflammatory and innate immune axis, COL7A1 is more closely linked to extracellular matrix and stromal phenotypes, WNT10B relates to cell-state regulation, and SCO2 suggests metabolic adaptation; together, these representative genes converge on the themes above, indicating that imaging features can capture molecular bases associated with tumor biological behavior.\u003c/p\u003e \u003cp\u003eBuilding on these findings, we further performed immune microenvironment analyses to provide cross-level validation of the \u0026ldquo;immune effector\u0026rdquo; axis. IPS showed that the high-risk group had higher antigen presentation\u0026ndash;related and effector immunity\u0026ndash;related components, while immune regulatory components also differed between groups. Consistently, MCPcounter inferred higher infiltration of cytotoxic lymphocytes, NK cells, and CD8 T cells in the high-risk group, together with differences in inflammation-related cellular components, suggesting that this stratification corresponds to a more complex immune ecosystem. Notably, prior imaging phenotype studies in ccRCC prognostic prediction have also reported concordance between imaging-derived stratification and immune infiltration composition, as well as immune-related transcriptional programs such as antigen presentation and interferon signaling, further supporting the potential of imaging phenotypes as indirect readouts of tumor immune status\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Collectively, these results provide a biologically coherent interpretive framework for imaging-based stratification of WHO/ISUP grading.\u003c/p\u003e \u003cp\u003eTo further evaluate the association between habitat features and pathway activity, we applied GSVA to quantify key biological processes and examined their relationships with habitat-derived spatial heterogeneity features. The GSVA results were concordant with the \u0026ldquo;immune effector and cell proliferation\u0026rdquo; axis highlighted by GSEA, and further mapped transcriptomic differences onto imaging-defined spatial phenotypes\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Notably, habitat features related to S1 (Sub-region 1) showed positive correlations with immune effector modules and cell cycle/replication modules, whereas S3 (Sub-region 3)\u0026ndash;related features more frequently exhibited negative correlations. This pattern suggests that distinct spatial subregions may correspond to different spectra of biological activity. Based on this \u0026ldquo;pathway activity and spatial subregion\u0026rdquo; mapping, we next used representative case visualizations to illustrate how the spatial configurations of S1 and S3 align with model outputs, thereby providing a more intuitive depiction of the spatial heterogeneity captured by the habitat framework.\u003c/p\u003e \u003cp\u003eThis study has several limitations. First, although we included multicenter cohorts and incorporated KiTS19 and TCGA for external validation, unavoidable heterogeneity exists across institutions and public databases in CT acquisition protocols, contrast-enhancement phases, reconstruction parameters, and clinical management pathways. Such variability may affect feature extraction and model stability. We mitigated this issue through standardized preprocessing and repeatability-based feature filtering; however, further validation in larger, prospective cohorts with more standardized multiparametric imaging protocols is still warranted. Second, WHO/ISUP grade was used as the primary endpoint, but differences in grade distribution and baseline characteristics across cohorts, together with the relatively limited clinical variables available in public datasets, may introduce potential confounding and influence both model performance and interpretability analyses. Third, prognostic evaluation was primarily based on Kaplan\u0026ndash;Meier curves and log-rank tests, and multivariable Cox regression analyses were not performed to adjust for tumor stage, treatment, and other key clinical factors. Therefore, the independent prognostic value of the model-derived risk stratification requires further assessment in cohorts with more complete clinical information.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis study proposes and validates a contrast-enhanced CT\u0026ndash;based habitat framework that captures intratumoral spatial heterogeneity and integrates score-level fusion to enable preoperative prediction of WHO/ISUP grade in clear cell renal cell carcinoma. The framework demonstrated stable predictive performance across multicenter cohorts as well as the public KiTS19 and TCGA cohorts. The model-derived risk score further enabled prognostic stratification of patients with ccRCC. Biological interpretability analyses showed that the risk groups were consistently associated with immune effector activity, proliferation programs, and cell-state\u0026ndash;related processes, with clear directionality at the subregion level, thereby providing a biologically coherent interpretive framework for imaging-based stratification. Overall, our findings support spatial heterogeneity\u0026ndash;aware imaging phenotypes as a potential noninvasive complement for ccRCC grading and risk assessment, and lay the groundwork for future prospective validation and deeper mechanistic investigations.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eccRCC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eClear cell renal cell carcinoma\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eISUP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eInternational Society of Urological Pathology\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eComputed tomography\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegion of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eITH\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntratumoral spatial heterogeneity\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eOS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOverall survival\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntraclass correlation coefficient\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eReceiver operating characteristic\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea under the ROC curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGSEA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGene set enrichment analysis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDEG\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDifferentially expressed gene\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZhe Jin, Siyi Chen, and Chen Jin contributed to methodology, investigation, formal analysis, and manuscript drafting. Zhe Jin also contributed to conceptualization, data curation, and visualization. Chen Jin was additionally responsible for software development and validation. Ying Ma, Yunshi Liang, Jingxuan Guo, Luyi Chen, Yingwen Liu, Xusheng Lin, Chuyi Huang, and Zhidan Zhong contributed to data curation and investigation, with Yunshi Liang, Yongxin Chen, Yuan Guo, and Wenjie Tang also contributing resources and/or validation. Weifeng Liu, Li Tian, and Xinqing Jiang contributed to study conception and design, supervision, project administration, and critical revision of the manuscript. Weifeng Liu and Xinqing Jiang also contributed to funding acquisition. All authors contributed to manuscript preparation, read, and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the National Natural Science Foundation of China (No. 82302314); Basic and Applied Basic Research Foundation of Guangdong Province (Nos. 2022A1515110792, 2023A1515220097, 2024A1515010653); Science and technology Projects in Guangzhou (Nos. 2024A03J1030, 2025A03J4162, 2025A03J4163); Guangdong Medical Research Fund (No. 202405300148274416); Clinical Research Hongmian Project of Guangzhou First People\u0026apos;s Hospital (No. HM2025062) .\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData sharing statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis multicenter retrospective study was conducted in accordance with the Declaration of Helsinki. The institutional review boards of Guangzhou First People\u0026rsquo;s Hospital and Sun Yat-sen University Cancer Center approved the study and waived the requirement for written informed consent due to its retrospective nature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of Interest\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSiegel RL, Kratzer TB, Wagle NS, Sung H, Jemal A. Cancer statistics, 2026. CA Cancer J Clin. 2026;76(1):e70043. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3322/caac.70043\u003c/span\u003e\u003cspan address=\"10.3322/caac.70043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBukavina L, Bensalah K, Bray F, Carlo M, Challacombe B, Karam JA, Kassouf W, Mitchell T, Montironi R, O'Brien T, et al. Epidemiology of Renal Cell Carcinoma: 2022 Update. Eur Urol. 2022;82(5):529\u0026ndash;42. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.eururo.2022.08.019\u003c/span\u003e\u003cspan address=\"10.1016/j.eururo.2022.08.019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDagher J, Delahunt B, Rioux-Leclercq N, Egevad L, Srigley JR, Coughlin G, Dunglinson N, Gianduzzo T, Kua B, Malone G, et al. Clear cell renal cell carcinoma: validation of World Health Organization/International Society of Urological Pathology grading. Histopathology. 2017;71(6):918\u0026ndash;25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/his.13311\u003c/span\u003e\u003cspan address=\"10.1111/his.13311\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohanty SK, Lobo A, Cheng L. The 2022 revision of the World Health Organization classification of tumors of the urinary system and male genital organs: advances and challenges. Hum Pathol. 2023;136:123\u0026ndash;43. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.humpath.2022.08.006\u003c/span\u003e\u003cspan address=\"10.1016/j.humpath.2022.08.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStewart GD, Klatte T, Cosmai L, Bex A, Lamb BW, Moch H, Sala E, Siva S, Porta C, Gallieni M. The multispeciality approach to the management of localised kidney cancer. Lancet. 2022;400(10351):523\u0026ndash;34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0140-6736(22)01059-5\u003c/span\u003e\u003cspan address=\"10.1016/S0140-6736(22)01059-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarris CR, Whitson JM, Meng MV. Under-grading of \u0026lt;\u0026thinsp;4 cm renal masses on renal biopsy. BJU Int. 2012;110(6):794\u0026ndash;7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1464-410X\u003c/span\u003e\u003cspan address=\"10.1111/j.1464-410X\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. .2012.10944.x From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi S, Zhou Z, Gao M, Liao Z, He K, Qu W, Li J, Kamel IR, Chu Q, Zhang Q, et al. Incremental value of automatically segmented perirenal adipose tissue for pathological grading of clear cell renal cell carcinoma: a multicenter cohort study. Int J Surg. 2024;110(7):4221\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/JS9.0000000000001358\u003c/span\u003e\u003cspan address=\"10.1097/JS9.0000000000001358\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShinagare AB, Davenport MS, Park H, Pedrosa I, Remer EM, Chandarana H, Doshi AM, Schieda N, Smith AD, Vikram R, et al. Lexicon for renal mass terms at CT and MRI: a consensus of the society of abdominal radiology disease-focused panel on renal cell carcinoma. Abdom Radiol (NY). 2021;46(2):703\u0026ndash;22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00261-020-02644-x\u003c/span\u003e\u003cspan address=\"10.1007/s00261-020-02644-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGillies RJ, Kinahan PE, Hricak H, Radiomics. Images Are More than Pictures, They Are Data. Radiology. 2016;278(2):563\u0026ndash;77. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.2015151169\u003c/span\u003e\u003cspan address=\"10.1148/radiol.2015151169\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuiot J, Vaidyanathan A, Deprez L, Zerka F, Danthine D, Frix AN, Lambin P, Bottari F, Tsoutzidis N, Miraglio B, et al. A review in radiomics: Making personalized medicine a reality via routine imaging. Med Res Rev. 2022;42(1):426\u0026ndash;40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/med.21846\u003c/span\u003e\u003cspan address=\"10.1002/med.21846\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao 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 (\u0026lt;\u0026thinsp;4 cm) clear cell renal cell carcinoma. BMC Cancer. 2023;23(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12885-023-11454-5\u003c/span\u003e\u003cspan address=\"10.1186/s12885-023-11454-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiao P, Wang B, Ni X, Lu Y, Yang R, Liu Y, Wang J, Mei H, Liu X, Weng X, et al. CT Urography-Based Radiomics to Predict ISUP Grading of Clear Cell Renal Cell Carcinoma. J Cancer. 2025;16(4):1118\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7150/jca.105173\u003c/span\u003e\u003cspan address=\"10.7150/jca.105173\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu C, Xi Y, Hu J, Li G, Wang X, Jiao X, Ding Z, Sun W. A predictive model for WHO/ISUP pathologic grading of renal clear cell carcinoma based on CT radiomics: a multicenter study. BMC Nephrol. 2025;26(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12882-025-04268-z\u003c/span\u003e\u003cspan address=\"10.1186/s12882-025-04268-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou H, Mao H, Dong D, Fang M, Gu D, Liu X, Xu M, Yang S, Zou J, Yin R, et al. Development and External Validation of Radiomics Approach for Nuclear Grading in Clear Cell Renal Cell Carcinoma. Ann Surg Oncol. 2020;27(10):4057\u0026ndash;65. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1245/s10434-020-08255-6\u003c/span\u003e\u003cspan address=\"10.1245/s10434-020-08255-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Su GH, Chen Y, Gu YJ, You C. Decoding Intratumoral Heterogeneity: Clinical Potential of Habitat Imaging based on Radiomics. Radiology. 2023;309(3):e232047. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.232047\u003c/span\u003e\u003cspan address=\"10.1148/radiol.232047\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTomaszewski MR, Gillies RJ. The Biological Meaning of Radiomic Features. Radiology. 2021;298(3):505\u0026ndash;16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.2021202553\u003c/span\u003e\u003cspan address=\"10.1148/radiol.2021202553\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi Z, Huang X, Cheng Z, Xu Z, Lin H, Liu C, Chen X, Liu C, Liang C, Lu C, et al. MRI-based Quantification of Intratumoral Heterogeneity for Predicting Treatment Response to Neoadjuvant Chemotherapy in Breast Cancer. Radiology. 2023;308(1):e222830. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.222830\u003c/span\u003e\u003cspan address=\"10.1148/radiol.222830\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang W, Jin C, Kong Q, Liu C, Chen S, Ding S, Liu B, Feng Z, Li Y, Dai Y, et al. Development and validation of an MRI spatiotemporal interaction model for early noninvasive prediction of neoadjuvant chemotherapy response in breast cancer: a multicentre study. EClinicalMedicine. 2025;85:103298. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.eclinm.2025.103298\u003c/span\u003e\u003cspan address=\"10.1016/j.eclinm.2025.103298\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM PubMed-not-MEDLINE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao R, Pang J, Lin P, Wen R, Wen D, Liang Y, Ma Z, Liang L, He Y, Yang H. Identification of clear cell renal cell carcinoma subtypes by integrating radiomics and transcriptomics. Heliyon. 2024;10(11):e31816. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.heliyon.2024.e31816\u003c/span\u003e\u003cspan address=\"10.1016/j.heliyon.2024.e31816\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM PubMed-not-MEDLINE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHartmann K, Sadee CY, Satwah I, Carrillo-Perez F, Gevaert O. Imaging genomics: data fusion in uncovering disease heritability. Trends Mol Med. 2023;29(2):141\u0026ndash;51. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoch H, Cubilla AL, Humphrey PA, Reuter VE, Ulbright TM. The 2016 WHO Classification of Tumours of the Urinary System and Male Genital Organs-Part A: Renal, Penile, and Testicular Tumours. Eur Urol. 2016;70(1):93\u0026ndash;105. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.eururo.2016.02.029\u003c/span\u003e\u003cspan address=\"10.1016/j.eururo.2016.02.029\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng Q, Mei H, Weng X, Yang R, Jiao P, Ni X, Yang X, Wu J, Fan J, Yuan J, et al. Artificial intelligence-based multimodal prediction for nuclear grading status and prognosis of clear cell renal cell carcinoma: a multicenter cohort study. Int J Surg. 2025;111(6):3722\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/JS9.0000000000002368\u003c/span\u003e\u003cspan address=\"10.1097/JS9.0000000000002368\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeller N, Sathianathen N, Kalapara A, Walczak E, Moore K, Kaluzniak H, Rosenberg J, Blake P, Rengel Z, Oestreich M. The KiTS19 Challenge Data: 300 Kidney Tumor Cases with Clinical Context, CT Semantic Segmentations, and Surgical Outcomes. \u003cem\u003earXiv e-prints\u003c/em\u003e 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo S, Wei R, Lu S, Lai S, Wu J, Wu Z, Pang X, Wei X, Jiang X, Zhen X, et al. Fuhrman nuclear grade prediction of clear cell renal cell carcinoma: influence of volume of interest delineation strategies on machine learning-based dynamic enhanced CT radiomics analysis. Eur Radiol. 2022;32(4):2340\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00330-021-08322-w\u003c/span\u003e\u003cspan address=\"10.1007/s00330-021-08322-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao 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 (\u0026lt;\u0026thinsp;4 cm) clear cell renal cell carcinoma. BMC Cancer. 2023;23(1):953. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12885-023-11454-5\u003c/span\u003e\u003cspan address=\"10.1186/s12885-023-11454-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiao P, Wang B, Ni X, Lu Y, Yang R, Liu Y, Wang J, Mei H, Liu X, Weng X, et al. CT Urography-Based Radiomics to Predict ISUP Grading of Clear Cell Renal Cell Carcinoma. J Cancer. 2025;16(4):1118\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7150/jca.105173\u003c/span\u003e\u003cspan address=\"10.7150/jca.105173\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM PubMed-not-MEDLINE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Z, Jiang H, Shan S, Wang X, Kou Q, Wang C, Jin P, Xu Y, Liu X, Zhang Y, et al. 2.5D Deep Learning-Based Prediction of Pathological Grading of Clear Cell Renal Cell Carcinoma Using Contrast-Enhanced CT: A Multicenter Study. Acad Radiol. 2025;32(10):5907\u0026ndash;16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.acra.2025.06.056\u003c/span\u003e\u003cspan address=\"10.1016/j.acra.2025.06.056\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang JL, Liu Q, Wang CL, Lang X, Zeng YX, Zhou DQ. CT-based subregional and peritumoral radiomics for predicting pathological T stage of clear cell renal cell carcinoma: an exploratory study of biological mechanisms. Insights Imaging. 2026;17(1):50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13244-026-02226-3\u003c/span\u003e\u003cspan address=\"10.1186/s13244-026-02226-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM PubMed-not-MEDLINE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang G, Nie P, Yan L, Zhang M, Wang Y, Zhao L, Li M, Xie F, Xie H, Li X, et al. The radiomics-based tumor heterogeneity adds incremental value to the existing prognostic models for predicting outcome in localized clear cell renal cell carcinoma: a multicenter study. Eur J Nucl Med Mol Imaging. 2022;49(8):2949\u0026ndash;59. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00259-022-05773-1\u003c/span\u003e\u003cspan address=\"10.1007/s00259-022-05773-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen S, Gao F, Guo T, Jiang L, Zhang N, Wang X, Zheng J. Deep learning-based multi-model prediction for disease-free survival status of patients with clear cell renal cell carcinoma after surgery: a multicenter cohort study. Int J Surg. 2024;110(5):2970\u0026ndash;7. 10.1097/JS9.0000000000001222 From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong S, Ge W, Qi X, Che X, Wang Q, Wu G. Computed tomography-based radiomics predicts prognostic and treatment-related levels of immune infiltration in the immune microenvironment of clear cell renal cell carcinoma. BMC Med Imaging. 2025;25(1):213. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12880-025-01749-3\u003c/span\u003e\u003cspan address=\"10.1186/s12880-025-01749-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiong Y, Yao L, Lin J, Yao J, Bai Q, Huang Y, Zhang X, Huang R, Wang R, Wang K, et al. Artificial intelligence links CT images to pathologic features and survival outcomes of renal masses. Nat Commun. 2025;16(1):1425. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-025-56784-z\u003c/span\u003e\u003cspan address=\"10.1038/s41467-025-56784-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchutte U, Bisht S, Heukamp LC, Kebschull M, Florin A, Haarmann J, Hoffmann P, Bendas G, Buettner R, Brossart P, et al. Hippo signaling mediates proliferation, invasiveness, and metastatic potential of clear cell renal cell carcinoma. Transl Oncol. 2014;7(2):309\u0026ndash;21. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tranon.2014.02.005\u003c/span\u003e\u003cspan address=\"10.1016/j.tranon.2014.02.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM PubMed-not-MEDLINE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan Z, Li Y, Zhang S, Wang X, Dou H, Yu X, Zhang Z, Yang S, Xiao M. Extracellular matrix remodeling in tumor progression and immune escape: from mechanisms to treatments. Mol Cancer. 2023;22(1):48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12943-023-01744-8\u003c/span\u003e\u003cspan address=\"10.1186/s12943-023-01744-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo Y, Gong B, Li Y, Lou J, Shan S, Zhang X, Lu Q, Zhu D, Sun Q, Miao L, et al. Radiomics-Based Unsupervised Clustering Identifies Subtypes Associated With Prognosis and Immune Microenvironment in Clear Cell Renal Cell Carcinoma: A Multicenter Study. Adv Sci (Weinh). 2025;12(34):e06165. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/advs.202506165\u003c/span\u003e\u003cspan address=\"10.1002/advs.202506165\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. From NLM Medline.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Clear cell renal cell carcinoma, WHO/ISUP grading, Computed tomography, Radiomics, Habitat Analysis","lastPublishedDoi":"10.21203/rs.3.rs-9303146/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9303146/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eTo develop and validate a preoperative contrast-enhanced CT\u0026ndash;based approach that leverages intratumoral spatial heterogeneity to predict WHO/ISUP grading and prognosis in clear cell renal cell carcinoma (ccRCC).\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis multicenter retrospective study included 704 patients with pathologically confirmed ccRCC (483 low-grade and 221 high-grade). Patients from SYUCC (n\u0026thinsp;=\u0026thinsp;308) were randomly split into training and internal validation sets at an 8:2 ratio. Two independent external validation cohorts were used, including GZFPH (n\u0026thinsp;=\u0026thinsp;106) and two public datasets, KiTS19 (n\u0026thinsp;=\u0026thinsp;142) and TCGA-KIRC (n\u0026thinsp;=\u0026thinsp;148). We constructed three models: a conventional whole-tumor radiomics model, an intratumoral habitat model capturing spatial heterogeneity, and a Combined model using score-level fusion of radiomic and habitat scores. Prognostic value was assessed in KiTS19 and TCGA using Kaplan\u0026ndash;Meier analysis with log-rank tests based on (i) ground-truth WHO/ISUP grade and (ii) model-derived risk groups. Biological interpretability was explored using differential expression and pathway enrichment analyses, GSVA-based pathway activity mapping to habitat subregions, and immune profiling (IPS and MCPcounter).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe Combined model achieved consistent performance across cohorts, with AUCs of 0.860 (95% CI: 0.813\u0026ndash;0.907) in SYUCC, 0.830 (0.813\u0026ndash;0.907) in GZFPH, 0.829 (0.756\u0026ndash;0.901) in KiTS19, and 0.750 (0.664\u0026ndash;0.830) in TCGA. Although AUC differences between the Combined and Habitat models were not statistically significant, the Combined model showed higher accuracy across all cohorts. Model-derived risk stratification significantly separated overall survival in both KiTS19 (p\u0026thinsp;=\u0026thinsp;0.017) and TCGA (p\u0026thinsp;=\u0026thinsp;0.0032). Transcriptomic analyses indicated coherent biological axes involving immune effector and proliferation programs versus differentiation and EMT/TGF-β\u0026ndash;related states, with GSVA revealing directionally distinct pathway associations for habitat subregions (S1 vs S3). Immune analyses further supported risk-group differences in antigen presentation and effector immunity components and in inferred immune cell infiltration.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eIntratumoral spatial heterogeneity on preoperative contrast-enhanced CT enables robust prediction of WHO/ISUP grade and prognostic stratification in ccRCC.\u003c/p\u003e","manuscriptTitle":"Preoperative CT Habitat Analysis for Predicting WHO/ISUP Grade in Clear Cell Renal Cell Carcinoma: A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-28 00:21:57","doi":"10.21203/rs.3.rs-9303146/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-19T10:49:37+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-11T05:09:36+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-10T12:22:36+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-10T12:21:34+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-04-02T12:19:57+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5e32886e-fb24-4d1e-95d3-8616ceae5eb7","owner":[],"postedDate":"April 28th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-28T00:21:57+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-28 00:21:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9303146","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9303146","identity":"rs-9303146","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}