Machine learning-based radiomics from multiparametric MRI for predicting aggressive pathology in clear cell renal cell carcinoma Authors: Jie Zhan, Lei Sun, Enming Cui, Zhitao Yang, Ting Zhang, Yanqing Yu, Panqi Xu, Jiayue Chen, Xin Zhen, Ruimeng Yang Published: 2025 ABSTRACT AbstractBackgroundClear cell renal cell carcinoma (ccRCC) exhibits significant biological heterogeneity, with aggressive forms demonstrating poor prognosis. Accurate preoperative discrimination between aggressive and indolent ccRCC is critical for individualized treatment but remains challenging. This study aimed to evaluate the performance of machine learning models based on multiparametric MRI radiomics for distinguishing aggressive from indolent ccRCC.MethodsThis retrospective study included 157 patients with pathologically confirmed ccRCC, comprising 114 indolent and 43 aggressive cases. Regions of interest (ROIs) were manually delineated on five MRI sequences: T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), as well as the corticomedullary, nephrographic, and excretory phases of contrast-enhanced fat-suppressed T1WI (CE-fsT1WI). Thirty-one feature combinations derived from the five sequences were input into 168 classification models (constructed using 8 classifiers and 21 feature selection methods). The performance of 5,208 models was compared, and the top-ranked features were analyzed.ResultsAggressive ccRCC showed significantly larger maximum tumor diameter compared with indolent tumors (8.3 [5.7–9.5] cm vs. 3.0 [2.2–4.2] cm, p < 0.05). Radiomic features derived from T2WI contributed most substantially to model performance relative to other MRI sequences, with the optimal classification model “RF + ICAP” achieving an area under the curve (AUC) of 0.960, accuracy of 86.1%, sensitivity of 86.4%, and specificity of 86.0%. Notably, the top 10 most predictive features from T2WI were predominantly shape-related features.ConclusionRadiomics features from renal T2WI demonstrated superior discriminative value compared with T1WI and contrast-enhanced T1WI in differentiating aggressive from indolent ccRCC. Through the integration of multiple classifiers and feature selection algorithms, the optimal classification model was identified, demonstrating the potential to distinguish aggressive ccRCC pathology.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12880-025-02046-9.Keywords: Clear cell renal cell carcinoma, Pathology, MRI, Machine learning, Classification Introduction Renal cell carcinoma (RCC) is the most common malignant tumor of the kidney, accounting for more than 90% of all solid renal cancers [1]. According to GLOBOCAN 2022 data [2], approximately 434,419 new cases of kidney cancer and more than 150,000 related deaths occurred worldwide, underscoring its considerable disease burden. Clear cell renal cell carcinoma (ccRCC), the predominant RCC subtype, comprises about 75%-80% of all RCC cases [3]. Owing to marked biological heterogeneity, ccRCC can present as either indolent or aggressive disease. Aggressive pathology is defined by nuclear grades 3–4, upstaging to pT3a, or the presence of sarcomatoid or dedifferentiated features [4, 5]. Compared to indolent ccRCC, aggressive subtypes are associated with worse overall survival and a significantly higher risk of metastasis, even at localized stages [6]. Accurate differentiation between aggressive and indolent ccRCC is therefore critical for prognosis and individualized treatment planning. Management of localized RCC has shifted toward a personalized approach, with clinical guidelines recommending treatment strategies tailored to tumor characteristics, comorbidities, and patient preferences [7]. Active surveillance may be appropriate for patients with indolent tumors, particularly those with significant comorbidities [8, 9]. In contrast, aggressive RCC generally requires early surgical intervention, such as partial or radical nephrectomy, to reduce the risk of progression and metastasis [10]. However, reliable preoperative identification of aggressive ccRCC remains challenging. Although percutaneous renal biopsy can provide histological confirmation, its invasive nature, potential complications (e.g., bleeding, infection), and sampling bias due to tumor heterogeneity limit its widespread use for preoperative risk stratification. Furthermore, since definitive histological diagnosis is obtained only after surgical resection, treatment decisions are inevitably delayed. Magnetic resonance imaging (MRI) offers a noninvasive alternative with superior soft-tissue contrast, enabling multi-parametric imaging that is valuable for renal mass detection and characterization [11, 12]. Previous studies have correlated MRI features with pathological characteristics of RCC, including histological subtypes [13, 14], nuclear grading [15–17], and tumor staging [17]. Nonetheless, few investigations have systematically explored the potential of MRI in predicting aggressive pathological features of RCC. Preliminary attempts have been made to address this gap. For example, Wang et al. [5] combined contrast-enhanced MRI parameters with clinical data to predict aggressive RCC pathology. However, such efforts relied primarily on two-dimensional measurements, neglecting the advantages of three-dimensional image analysis, including comprehensive feature extraction, morphological assessment, and texture characterization. Consequently, critical biological information embedded within the full three-dimensional tumor architecture remains underutilized. Recent advances in radiomics and artificial intelligence (AI) offer promising alternatives for non-invasive prediction of tumor biology. Radiomics enables extraction of high-dimensional quantitative features—including intensity, shape, size, and texture—that can characterize tumor phenotype and microenvironment beyond conventional clinical or pathological assessments [18]. Multiparametric MRI-based radiomics has shown considerable potential in differentiating benign and malignant renal tumors [19, 20], predicting nuclear grading [21–23], and identifying histological subtypes [24]. Nevertheless, studies specifically focusing on the preoperative prediction of tumor aggressiveness remain limited. Recently, Du et al. [25] proposed a convolutional neural network-based fusion model integrating multiparametric diffusion-weighted imaging (DWI) with clinical data to predict RCC aggressiveness preoperatively, with promising results. However, their model excluded conventional MRI sequences such as T1WI, T2WI, and contrast-enhanced imaging, thereby limiting its ability to capture key morphological features critical for accurate staging and aggressiveness assessment in RCC. Given that ccRCC is the predominant histological subtype of RCC, early and accurate risk stratification of its aggressiveness is of paramount importance in clinical management. This study aims to systematically evaluate and compare the diagnostic performance of different MRI sequences and combinations of radiomics features for differentiating aggressive from indolent ccRCC. In addition, multiple feature selection methods and machine learning classifiers will be explored to establish an optimal predictive model, with the ultimate goal of exploring its potential as a tool for preoperative risk assessment following future validation. Materials and methods Patient cohort This retrospective study included patients with pathologically confirmed ccRCC from three institutions: the First Affiliated Hospital of Nanchang University (Center 1), the Central Hospital of Jiangmen City (Center 2), and The Cancer Imaging Archive (TCIA, Center 3). The ethics committees of the First Affiliated Hospital of Nanchang University and the Central Hospital of Jiangmen City waived the requirement for informed consent. Informed consent was also waived for the use of TCIA data. Inclusion criteria were as follows: (1) histopathological confirmation of ccRCC after partial or radical nephrectomy; (2) MRI examination performed within two weeks prior to surgery; and (3) availability of complete clinical, pathological, and imaging data. Exclusion criteria were: (1) poor MRI image quality due to motion artifacts or other technical issues; (2) history of interventional therapy or recurrent ccRCC; and (3) coexistence of other malignancies. Ultimately, a total of 157 patients were included in the study. Image acquisition All MRI examinations at Center 1 were performed on a 3.0-T scanner (Magnetom Trio, Siemens Healthcare, Erlangen, Germany) equipped with a 32-channel body coil. At Center 2, MRI examinations were conducted using either a 1.5-T scanner (Achieva, Philips Healthcare) or a 3.0-T scanner (Ingenia, Philips Healthcare), both equipped with a 32-channel body coil. The data for Center 3 were obtained from TCIA and were acquired on scanners from multiple manufacturers, including Siemens (Magnetom Vision Plus, Avanto, Triotim), GE Medical Systems (Genesis Signa), and Philips (Achieva), with field strengths ranging from 1.0 T to 3.0 T. The sequences used for analysis included: (1) an axial T1-weighted imaging (T1WI); (2) an axial T2-weighted imaging (T2WI) with or without fat suppression; and (3) multiphasic contrast-enhanced fat-suppressed T1-weighted imaging (CE-fsT1WI), including the corticomedullary phases (CMP), nephrographic phases (NP), and excretory phases (EP). The CE-fsT1WI was performed by intravenous injection of a gadolinium-based contrast agent at a dose of 0.2 ml/kg, with images acquired at 20–30 s, 70–80 s, and 180 s post-injection. The detailed parameters for all centers, including those for the TCIA dataset (Center 3) which exhibit inherent variability due to its multi-institutional nature, are summarized in Supplementary Table 1. Histopathological analysis Pathological diagnosis was confirmed by two experienced urological pathologists after surgery. The tumors were graded according to the World Health Organization and International Society of Urological Pathology (WHO/ISUP) grading system [26] and were staged according to the American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) staging systems [27]. Aggressive pathology in ccRCC was defined as the presence of any of the following [4, 5]: (1) nuclear grade 3–4; (2) pathological stage ≥ pT3a; or (3) presence of sarcomatoid or dedifferentiated features. Image preprocessing All MRI data were stored in Digital Imaging and Communications in Medicine (DICOM) format and anonymized. To ensure data quality and minimize technical variances arising from the multi-center and multi-scanner nature of this study, a standardized image preprocessing pipeline was implemented prior to feature extraction. This pipeline involved key steps to address spatial and intensity inconsistencies. Specifically, all MR images underwent spatial normalization, where they were resampled to a uniform isotropic voxel space using B-spline interpolation and co-registered across sequences for each patient to correct for positional discrepancies. Furthermore, to mitigate scanner-related biases in absolute intensity values, intensity normalization was performed by applying Z-score standardization, scaling the intensity values of each sequence to a distribution with a mean of zero and a standard deviation of one. Feature extraction and representation Image segmentation was performed manually by a radiologist (J Zhan, 5 years of experience) using ITK-SNAP software (http://www.itksnap.org) without knowledge of the pathological results. Regions of interest (ROIs) were delineated on each MRI sequence and phase (refer to Fig. 1), and volumes of interest (VOIs) were generated for each patient. A senior radiologist (R.M Yang, 15 years of experience) reviewed and refined all segmentations. Radiomic feature extraction was performed using Pyradiomics (version 3.0.1) [28]. Only features with an intraclass correlation coefficient (ICC) >0.75 and reproducibility across all five MR sequences were retained. A total of 107 candidate features were extracted (refer to Table 1), categorized as: (1) shape features, (2) first-order statistics (histogram analysis), and (3) second-order statistics (gray-level distribution of the image, commonly called “texture features”). Detailed feature definitions are available in the Pyradiomics documentation [28]. Representative MR images from T1WI, T2WI, CMP, NP, and EP in a patient with pathologically proven ccRCC. The lower panel shows magnified views corresponding to the regions highlighted by rectangles in the upper panel. T1WI, T1-weighted imaging; T2WI, T2-weighted imaging; CMP, corticomedullary phase; NP, nephrographic phase; EP, excretory phase; ccRCC, clear cell renal cell carcinoma Table 1Extracted candidate featuresShape features(n = 14)Elongation, Flatness, Least Axis Length, Major Axis Length, Maximum 2D Diameter Column, Maximum 2D Diameter Row, Maximum 2D Diameter Slice, Maximum 3D Diameter, Mesh Volume, Minor Axis Length, Sphericity, Surface Area, Surface Volume Ratio, Voxel VolumeFirst-order features(n = 18)10 percentile, 90 percentile, energy, entropy, interquartile range, kurtosis, minimum, range, mean absolute deviation, mean, median, range, robust mean absolute deviation, root mean squared, skewness, total energy, uniformity, varianceTexture features(n = 75)GLCM (n = 24), GLSZM (n = 16), GLRLM (n = 16), NGTDM (n = 5), GLDM (n = 14) Extracted candidate features Shape features (n = 14) First-order features (n = 18) Texture features (n = 75) Radiomic features were extracted from five MRI sequences: T1WI (), T2WI (), and the CMP (), NP (), and EP () of CE-fsT1WI. Each sequence yielded 107 features, which were systematically grouped into five categories to evaluate multi-sequence synergy: group 1 comprised features from individual sequences (–, each containing 107 features); group 2 included pairwise combinations (e.g., for T1WI + T2WI; 10 sets, 214 features each); group 3 consisted of triple-sequence combinations (e.g., ; 10 sets, 321 features each); group 4 involved four-sequence combinations (e.g., ; 5 sets, 428 features each); and group 5 combined all five sequences (; 535 features). In total, 31 (5 + 10 + 10 + 5 + 1) distinct feature sets were generated and served as inputs for discrimination models to assess their diagnostic power. Construction and comparison of predictive models In this study, models were constructed by integrating feature selection strategies with classifiers, and their performance was evaluated using fivefold cross-validation. In each fold, a specific feature selection method was first applied to select the optimal feature subset from the complete feature set. Subsequently, the selected features were fed into a classifier to build the discrimination model. A total of 168 different models were evaluated, incorporating 21 feature selection methods and 8 classifiers. The specific feature selection methods and classifiers used are listed in Supplementary Tables 2, and the explanations of the corresponding acronyms are provided in Supplementary Table 3. To ensure unbiased comparison, the number of features selected by each method was standardized to 20, based on the relationship between sample size and the number of features. To address the data imbalance in the patient cohort, the synthetic minority over-sampling technique (SMOTE) [29] was employed to oversample the aggressive ccRCC group by generating synthetic samples from the minority class. SMOTE has been shown to prevent overfitting and improve model generalizability [29, 30]. Specifically, within each fold of the fivefold cross-validation, SMOTE was applied exclusively to the training set to generate synthetic samples, while the validation set remained unaltered to prevent data leakage and ensure an unbiased evaluation of model performance. In total, 5,208 model combinations (31 feature sets × 168 models) were tested. Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE). The overall workflow of this study is shown in Fig. 2. The overall workflow of this study Statistical analysis Statistical analysis was conducted using SPSS 25.0 software (SPSS Inc.). Continuous variables were presented as mean ± standard deviation (SD) for normally distributed data or median (interquartile range, IQR) for non-normally distributed data. Categorical variables were expressed as frequencies and percentages. Normality was assessed using the Shapiro-Wilk test. Group comparisons were conducted using the Chi-square test for categorical variables, Student’s t test for normally distributed continuous variables, and the Mann-Whitney U test for non-normally distributed continuous variables. Comparisons of discriminative performance across the 31 feature categories were performed with the Kruskal-Wallis test, followed by Bonferroni-adjusted pairwise post hoc tests. A p value of < 0.05 was considered statistically significant. Results Patient cohort A total of 157 patients (93 males, 64 females) were included in the study after applying the inclusion and exclusion criteria. Patient characteristics are summarized in Table 2. No significant differences in sex distribution were observed between the indolent and aggressive groups (p > 0.05). However, patients in the aggressive group were significantly older and had larger maximum tumor diameters compared with those in the indolent group (p < 0.001). Table 2Demographics and characteristics of the patient cohortCharacteristicIndolent (n = 114)Aggressive (n = 43)p valuePatient age (median (IQR), year)51.0 (41.8, 64.3)60.0 (49.0, 66.0)0.007a*Sex0.357b Male (n/%)65 (57.0%)28 (65.1%) Female (n/%)49 (43.0%)15 (34.9%)Lesion diameter (median (IQR), cm)3.0 (2.2, 4.2)8.3 (5.7, 9.5)< 0.001a*Data were described as median (IQR) or frequency (%). a Mann-Whitney U test; b Chi-square test. * A p value less than 0.05 was considered statistically significant Demographics and characteristics of the patient cohort Data were described as median (IQR) or frequency (%). a Mann-Whitney U test; b Chi-square test. * A p value less than 0.05 was considered statistically significant Comparisons across feature sets Thirty-one feature sets were evaluated across 168 models, and their performances were compared. For each model, the best-performing feature category was identified and visualized in a heatmap (Fig. 3). The frequency with which each feature category ranked first is summarized in Table 3. T2WI-derived features achieved the highest performance most frequently (31 times), followed by composite feature sets involving T2WI, including T2WI-NP-EP (14 times), T2WI-NP (12 times), and T2WI-CMP-NP (11 times). These findings underscore the dominant role of T2WI-derived features in distinguishing aggressive from indolent ccRCC. Each cell in the color map represents one of 168 discriminative models, each built with a unique combination of classifier and feature-selection method. The color denotes the feature type that yielded the highest AUC among the 31 possibilities for that model. AUC, area under the receiver operating characteristic curve Table 3Top discriminative performance of 31 feature types across 168 modelsGroupMRAUC (ACC/SEN/SPE)ClassifierFeature selection methodNumber of timesbeing rankedas the bestfeature typeGroup 1 0.943 (0.828/0.842/0.824)LRMIM1 0.960 (0.861/0.864/0.860) RF ICAP 31 0.943 (0.854/0.822/0.868)RFFSCO10 0.944 (0.835/0.814/0.842)RFJMI4 0.941 (0.854/0.842/0.859)LRLSCO3Group 2 0.946 (0.835/0.911/0.807)LRGIND7 0.944 (0.835/0.775/0.860)LRLS_L214 0.942 (0.835/0.842/0.833)LRICAP0 0.946 (0.829/0.842/0.824)LRICAP2 0.959 (0.835/0.839/0.834) RF ICAP 8 0.956 (0.854/0.794/0.877)RFMIM12 0.956 (0.854/0.817/0.869)RFMIM6 0.946 (0.834/0.842/0.832)RFMIFS7 0.941 (0.860/0.792/0.886)RFCIFE4 0.938 (0.854/0.839/0.860)RFDISR4Group 3 0.944 (0.796/0.864/0.772)LRGIND4 0.944 (0.816/0.864/0.798)NBTSCO5 0.950 (0.848/0.842/0.851)LRJMI6 0.942 (0.835/0.819/0.842)LRCMIM0 0.943 (0.854/0.817/0.868)RFTSCO0 0.941 (0.816/0.864/0.797)LRMIM1 0.959 (0.848/0.864/0.842) RF ICAP 11 0.958 (0.854/0.819/0.869)RFMIM6 0.966 (0.867/0.864/0.869) RF CMIM 14 0.948 (0.854/0.842/0.859)LRMCFS3Group 4 0.944 (0.828/0.842/0.824)LRCIFE8 0.948 (0.841/0.844/0.842)LRJMI0 0.942 (0.815/0.908/0.780)NBTSCO2 0.937 (0.822/0.819/0.824)LRMIM0 0.956 (0.867/0.842/0.877)RFICAP3Group 5 0.947 (0.842/0.842/0.842)LRCMIM2The values in italics indicate the best AUC (ACC/SEN/SPE) of all Groups 1 ~ 5, the values in bold denote the top-three performing models overall Top discriminative performance of 31 feature types across 168 models The values in italics indicate the best AUC (ACC/SEN/SPE) of all Groups 1 ~ 5, the values in bold denote the top-three performing models overall This conclusion was further supported by ROC curve analysis (see Supplementary Fig. 1), generated using the classifier and feature selection method that achieved the highest discriminative performance. The best AUC (0.966) was obtained with a combination of T2WI, NP, and EP sequences, using the “RF” classifier with “CMIM” feature selection (Table 3). Figure 4 illustrates the distribution of AUC values across all 168 models. Overall, feature sets incorporating T2WI consistently outperformed those without T2WI. Boxplots of the AUC distributions achieved by the 168 discriminative models for each of the 31 feature types. The boxes run from the 25th to 75th percentile; the two ends of the whiskers represent the 5% and 95% percentiles of the data; the horizontal line and the square in the box represent the median and mean values, respectively. The circles represent outliers. The letters above each box indicate whether a statistically significant difference (by the Kruskal-Wallis test with the Bonferroni correction) exists between any two feature types. No common letters indicate that the two feature types are significantly different. AUC, area under the receiver operating characteristic curve Dominant role of T2WI: analysis of Top-Ranked predictive features The AUC performance of all 168 model configurations using is presented as a heatmap in Fig. 5 (see Supplementary Fig. 2 for corresponding heatmaps of ACC, SEN, and SPE). For models with an AUC > 0.85, the frequency of each feature appearing among the top 20 across fivefold cross-validation was calculated and visualized as a histogram (Fig. 6). The top 10 most frequently selected features in are marked as blue in Fig. 6 and summarized in Table 4, all of which are shape features. All ten features demonstrated high statistical significance, with p-values < 10⁻¹². AUCs of all the 168 discriminative models using as feature input. AUC, area under the receiver operating characteristic curve Histogram representing the number of times of each feature in being selected into the top 20 features in the fivefold cross-validation of all discriminative models with AUC > 0.85. AUC, area under the receiver operating characteristic curve Table 4Top10 most frequently selected features in and group differenceFeature categoryTop-ranked featuresAggressiveIndonentp-valueM†1(< M | >M)†2Shape-basedLeastAxisLength50.88 (17.88, 105.29)22.12 (0, 58.60)< 10− 15,b36.5Aggressive (30.2%|69.8%)Indonent (90.4%|9.6%)MajorAxisLength80.50 (37.66, 182.28)33.35 (9.09, 97.01)< 10− 17,b56.93Aggressive (25.6%|74.4%)Indonent (93.0%|7.0%)Maximum2DDiameterColumn88.70 (43.27, 208.81)37.03 (9.43, 109.18)< 10− 17,b62.86Aggressive(30.2%|69.8%)Indonent (92.1%|7.9%)Maximum2DDiameterRow88.49 (36.24, 206.16)36.67 (10.00, 102.18)< 10− 17,b62.58Aggressive (23.3%|76.7%)Indonent (93.9%|6.1%)Maximum2DDiameterSlice87.62 (40.00, 302.77)34.14 (8.54, 87.24)< 10− 18,b60.88Aggressive (23.3%|76.7%)Indonent (95.6%|4.4%)Maximum3DDiameter103.14 (45.00, 304.56)40.14 (11.36, 109.90)< 10− 17,b71.64Aggressive (25.6%|74.4%)Indonent (94.7%|5.3%)MinorAxisLength62.28 (28.64, 121.69)26.47 (6.48, 67.06)< 10− 17,b44.37Aggressive(25.6%|74.4%)Indonent (93.0%|7.0%)Sphericity0.69 (0.52, 0.81)0.77 (0.65, 0.85)< 10− 12,b0.73Aggressive (81.4%|18.6%)Indonent (17.5%|82.5%)SurfaceArea25114.86 (4642.36, 102652.30)4201.54 (239.53, 26279.12)< 10− 17,b14658.2Aggressive (34.9%|65.1%)Indonent (96.5%|3.5%)SurfaceVolumeRatio0.14 (0.07, 0.25)0.32 (0.10, 0.94)< 10− 16,bM: 0.23Aggressive (93.0%|7.0%)Indonent (28.1%|71.9%)†1 M, the mean of the mean feature value of the Aggressive group and the mean feature value of the Indonent group†2 Percentage of patients in Aggressive and Indonent groups with feature value less than or larger than the “M” valuea Independent t testb Mann-Whitney U test Top10 most frequently selected features in and group difference †1 M, the mean of the mean feature value of the Aggressive group and the mean feature value of the Indonent group †2 Percentage of patients in Aggressive and Indonent groups with feature value less than or larger than the “M” value a Independent t test b Mann-Whitney U test When the median value of each feature between the two groups (denoted as “M” in Table 4) was used as the threshold for classification, three shape features—Maximum2DDiameterRow, Maximum2DDiameterSlice, and Sphericity—exhibited particularly strong discriminative power. Specifically, for Maximum2DDiameterRow and Maximum2DDiameterSlice, 76.7% of aggressive ccRCC cases had feature values above the threshold M, whereas 93.9% and 95.6% of indolent cases, respectively, had values below M. Conversely, for Sphericity, 82.5% of indolent cases had values above M, while 81.4% of aggressive cases had values below M. Overall, these three morphological features each achieved a unilateral classification accuracy exceeding 75% at threshold M, providing a robust basis for reliably discriminating between indolent and aggressive ccRCC. Discussion The principal finding of this study is that T2WI-derived radiomic features are the most reliable for differentiating aggressive from indolent ccRCC. Among all models tested, the combination of “RF” with “CMIM” achieved the highest AUC (0.966), closely followed by “RF” with “ICAP” (AUC = 0.960). The top-performing model integrated T2WI, NP, and EP sequences, while the second model relied solely on T2WI, underscoring the pivotal role of T2WI-derived features. Previous studies have investigated the potential of MRI-based radiomics for similar clinical challenges. For instance, Purkayastha et al. [31] employed 13 feature selection methods and 10 machine learning classifiers to distinguish low- and high-grade RCC using routine MRI radiomics features. Their findings indicated that an automated machine learning pipeline was comparable or superior to a manually optimized one for predicting low-grade RCC during external validation. This highlights the significant potential of MRI-based radiomics in grading RCC, consistent with our findings, in which MRI-based characteristics were pivotal in distinguishing aggressive from indolent ccRCC. Similarly, Yang et al. [21] investigated the predictive value of radiomics features extracted from fat-suppressed T2WI, DWI, and contrast-enhanced T1WI sequences for ccRCC nuclear grading using several machine learning classifiers. They reported that the multilayer perceptron classifier provided the best performance, further supporting the potential of radiomics in predicting the WHO/ISUP nuclear grades. However, these prior studies were limited by reliance on single-sequence features, restricted classifier selection, and relatively small sample sizes, which constrain generalizability. In contrast, our study conducted a more comprehensive evaluation in a larger patient cohort, systematically combining eight robust classifiers and 21 feature selection methods to generate 168 discrimination models, enabling a broader and more rigorous assessment. Notably, our conclusions are consistent with a recent meta-analysis evaluating MRI-based radiomics for grading clear cell renal cell carcinoma (ccRCC). Broomand Lomer et al. reported that radiomics-only models achieved a pooled area under the curve (AUC) of 0.84, confirming the robustness of this approach and supporting the clinical relevance of our research direction [32]. Their analysis further showed that combined models integrating clinical and radiological semantic features achieved superior performance (pooled AUC = 0.90). This finding suggests that incorporating readily available indicators—such as tumor size, patient age, and the presence of necrosis—alongside radiomic features may enhance clinical applicability. In addition, subgroup analysis revealed that segmentation performed by a single operator improved model specificity, consistent with our use of a fixed senior radiologist to review and refine all segmentations. Overall, this meta-analysis provides external validation for our findings and delineates a clear pathway toward developing more robust and clinically translatable combined models. Clear cell renal cell carcinoma (ccRCC) is a heterogeneous tumor with varying degrees of necrosis, hemorrhage, and cystic changes, which are closely associated with tumor grade and prognosis. Low-grade ccRCC often presents with cystic features, while high-grade ccRCC frequently exhibits intratumoral necrosis. T2WI plays a crucial role in non-invasively assessing ccRCC, capturing these internal features and facilitating differentiation of aggressive tumors. For example, Takeuchi et al. [33] discovered that sarcomatoid differentiation in ccRCC can be distinguished by its hypointense features on T2WI. In addition, T2WI-derived quantitative metrics, such as T2 values and shape features, have been linked to tumor grade and stage. Li et al. [34] demonstrated that higher-grade and more advanced ccRCCs are associated with lower T2 values. Our results are consistent with these reports, indicating that T2WI-derived radiomic signatures serve as robust predictors of aggressive ccRCC. We propose that the superior performance of T2WI-derived features can be attributed to two key factors. First, T2WI is highly sensitive to variations in tissue free water content, enabling clear differentiation between solid and necrotic or cystic regions without the need for exogenous contrast agents. Necrosis or cystic degeneration—a hallmark of high-grade ccRCC—appears as marked hyperintensity on T2WI, providing a direct imaging–pathology correspondence that offers a biologically grounded and stable basis for radiomic feature extraction. Second, T2WI image quality is not influenced by individual hemodynamic variability or differences in contrast injection timing, ensuring high reproducibility across multicenter and multibatch studies. This reproducibility represents a distinct advantage over CE-fsT1WI, which remains susceptible to such technical fluctuations. By implementing a standardized radiomics pipeline, we extracted comprehensive textural and morphological features from T2WI, achieving high accuracy in differentiating aggressive from indolent tumors and reinforcing the potential of T2WI radiomics in noninvasive ccRCC phenotyping. A notable strength of this study lies in the identification of specific T2WI-derived shape features—Maximum2DDiameterRow, Maximum2DDiameterSlice, and Sphericity—as top three reliable predictors of aggressive ccRCC. These descriptors, which capture tumor size and structural irregularity, have long been implicated in tumor aggressiveness [35–38]. For instance, Bhindi et al. [35] reported a significant association between tumor size and pathological aggressiveness, with larger tumors more likely to exhibit high-grade histological features. Similarly, Shimada et al. [36] demonstrated that tumor shape irregularity, quantified by the radiomic feature “SHAPE_Sphericity”, had a strong predictive value for pT3a upstaging in localized cT1b-cT2 RCCs. More recently, Yuan et al. [39] developed a CT-based radiomics model that reliably identified pT3a invasion in RCC, achieving AUCs >0.84 across both internal and external validations, with “Sphericity” emerging as a key predictive feature. Consistent with these findings, our cohort showed that T2WI-derived shape features provided strong discriminatory power, underscoring their potential as reliable imaging biomarkers for stratifying ccRCC aggressiveness. Notably, the utility of T2WI-based radiomics is not confined to renal cell carcinoma. Similar machine learning-based workflows have demonstrated strong predictive performance for ISUP grade group differentiation in prostate MRI, another major genitourinary malignancy [40]. This evidence underscores the cross-cancer applicability of T2WI-derived features as imaging biomarkers across genitourinary cancers. ​In this study, our results indicate that sphericity has strong potential in predicting the aggressiveness of ccRCC. The underlying mechanisms can be interpreted from multiple perspectives. At the cellular level, malignant cells exhibit autonomous and irregular growth patterns, resulting in disruption of normal anatomic or histological structures [41, 42]. At the pathological level, previous studies showed that tumor irregularity was an indicator of intrinsic tumor aggressiveness [43, 44]. In addition, previous studies have reported that the renal parenchymal infiltration or micronodular satellitosis (RPI/MNS) phenotype is closely associated with highly aggressive ccRCC [45]. Therefore, tumor irregularity may be intrinsically linked to the RPI/MNS pattern, providing a biological rationale for the predictive value of sphericity. However, the retrospective design of the present study precludes direct histological validation of this association. Future studies integrating multiparametric imaging with detailed pathological correlation are warranted to elucidate the biological mechanisms underlying this morphological feature. Interestingly, although prior studies suggested sex as a predictor of tumor aggressiveness, we observed no statistically significant difference between males and females in our cohort. Reports from Bhindi et al. [35] and analyses of the Surveillance Epidemiology and End Results (SEER) database [46] indicated that male might be associated with higher-grade or more aggressive RCC. The lack of a similar association in our study could be due to the relatively limited sample size or specific characteristics of our patient population. Moreover, the biological mechanisms underlying potential sex-related differences in RCC behavior remain poorly understood and require further investigation. In contrast, our findings revealed a significant association between age and tumor aggressiveness, with older patients more likely to present with aggressive ccRCC. This result is in line with previous studies by Liao et al. [47] and Taccoen et al. [48], which also demonstrated a positive correlation between advancing age, tumor grade, and adverse pathological features. These observations highlight the prognostic relevance of age in ccRCC and underscore the need for age-stratified strategies in future clinical risk models and treatment planning. This study has several limitations. First, the retrospective design and the relatively small absolute number of aggressive cases (n = 43) may constrain the external validity and generalizability of our findings. Although fivefold cross-validation and the SMOTE were employed to mitigate overfitting, the risk of model overfitting cannot be entirely ruled out due to the limited sample size. Therefore, future independent external validation in larger cohorts is warranted to verify the robustness of our model. Second, the study cohort primarily consisted of patients ineligible for contrast-enhanced CT, which could introduce selection bias. Prospective studies with larger and more diverse populations are thus warranted. Third, DWI, which provides complementary information on tumor cellularity and microstructure, was not included. Incorporating DWI in future investigations may improve model performance and enhance generalizability across clinical settings. Fourth, multi-center variability in MRI acquisition protocols and the use of different scanner types may have introduced technical inconsistencies. The adoption of standardized imaging protocols across institutions would help mitigate such variability and improve the reproducibility of radiomics analyses. Conclusion In conclusion, this study demonstrates that MRI-based radiomics, particularly T2WI-derived shape features, can reliably distinguish aggressive from indolent ccRCC. By leveraging advanced machine learning methods and rigorous feature selection strategies, our models achieved high discriminatory performance, with T2WI-based descriptors of tumor size and morphology emerging as robust imaging biomarkers. These findings underscore the potential value of T2WI radiomics for the noninvasive assessment of ccRCC aggressiveness. Future prospective, multicenter studies incorporating additional MRI sequences (e.g., DWI) are warranted to validate and refine these models for broader clinical translation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Abbreviations Accuracy Artificial intelligence American Joint Committee on Cancer Area under the curve Clear cell renal cell carcinoma Contrast-enhanced fat-suppressed T1-weighted imaging Corticomedullary phases Digital Imaging and Communications in Medicine Diffusion-weighted imaging Excretory phases Intraclass correlation coefficient Magnetic resonance imaging Nephrographic phases Renal cell carcinoma Region of interest Renal parenchymal infiltration or micronodular satellitosis Sensitivity Synthetic minority over-sampling technique Specificity T1-weighted imaging T2-weighted imaging The Cancer Imaging Archive Tumor, node, metastasis Volume of interest World Health Organization and International Society of Urological Pathology Author contributions JZ, LS, EMC, ZTY, TZ, YQY, PQX, JYC, and XZ were in charge of data acquisition, analysis and interpretation. JZ, LS, EMC, ZTY, TZ were responsible for manuscript drafting. XZ, and RMY critically reviewed and revised the manuscript and supervised the study. JZ, LS, EMC, XZ, and RMY made a contribution to study conception and design. All authors contributed to the manuscript preparation, read and approved the final text. Funding This work was supported by the National Natural Science Foundations of China (82371908, 82572381, 62262029), the GuangDong Natural Science Foundation (2024A1515012177, 2024A1515012100), the Jiangxi Provincial Natural Science Foundation Project (20242BAB25546), the Jiangxi Clinical Research Center for Medical Imaging (20223BCG74001) and Jiangxi Province Key Laboratory for Precision Pathology and Intelligent Diagnosis (2024SSY06281). Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate The study has been performed in accordance with the Declaration of Helsinki. The study was approved by the Ethics Committee of the First Affiliated Hospital of Nanchang University (Approval No. IIT2024814) and the Central Hospital of Jiangmen City (Approval No. IIT20235). Given the retrospective nature of the study, the requirement for informed consent was waived by the ethics committees. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Jie Zhan, Lei Sun and Enming Cui contributed equally to this work as co-first authors. Contributor Information Xin Zhen, Email: xinzhen@smu.edu.cn. Ruimeng Yang, Email: eyruimengyang@scut.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.