This study analyzed data from 372 individuals diagnosed with urothelial carcinoma at our hospital between January 2015 and October 2020. Of these, 182 had low-grade and 190 had high-grade carcinoma. Inclusion criteria required a postoperative pathological diagnosis, complete clinical records, and a contrast-enhanced CT scan showing a bladder tumor of at least 1 cm in diameter. Exclusions included prior neoadjuvant therapy, inadequate bladder filling, tumors <1 cm, organ insufficiencies, inflammatory bladder disease, psychiatric disorders, and other malignancies. Patients were split into training and validation groups in a 7:3 ratio, as shown in Figure 1 .

Collected clinical data included height, age, sex, urinary pH, weight, presence of metabolic syndrome (MS), history of hypertension, urinary tract infections, and body mass index (BMI). It also included hematuria, urinary protein, high-density lipoprotein cholesterol (HDL-C), and tumor location, number, and size (maximum diameter). Additional data included diabetes mellitus (DM), triglyceride levels (TG), pathological grade, past medical history, and family history. Imaging data comprised thin-slice contrast-enhanced CT images of the venous and arterial phases, taken within a month prior to surgery, clearly showing bladder tumors without artifacts.

All participants underwent abdominal and pelvic (bladder) arteriovenous CT scans using a Siemens 64-slice spiral CT scanner. Reconstructed images were uploaded in uncompressed DICOM format to the Picture Archiving and Communication System (PACS). Arterial and venous phase thin-layer image data were then downloaded from the PACS system.

Arterial and venous phase images were fused using the 3D Slicer image processing software. Two radiology residents, each with over three years of experience in urinary system imaging interpretation, independently delineated the ROI using the Threshold and Sphere Brush functions, without prior knowledge of pathological results. Semi-automatic, continuous-slice ROI delineation was performed to encompass the entire bladder tumor lesion. To ensure morphological consistency, contours were smoothed after segmentation. The ROI delineation process is illustrated in Figure 2 . Furthermore, to encompass the entire bladder tumor as accurately as possible, the delineation line is maintained approximately 1–2 mm from the tumor’s edge. In cases of multiple foci, only the largest focus is included in the study to account for individual variability.

Radiomic feature extraction was carried out with the Radiomics module in 3D Slicer, which calls PyRadiomics v3.0 and fully implements Image Biomarker Standardization Initiative (IBSI) definitions. To minimize inter-scanner variability, all CT volumes were (i) resampled to isotropic 1 × 1 × 1 mm³ voxels, (ii) z-score intensity-normalized, and (iii) further harmonized across scanners using the ComBat algorithm. IBSI-compliant features were then generated, including first-order statistics, shape descriptors (Shape2D, Shape3D), and higher-order texture matrices—gray-level co-occurrence (GLCM), run-length (GLRLM), size-zone (GLSZM), dependence (GLDM), and neighboring gray-tone difference (NGTDM). Multiscale information was captured with log-sigma filters (σ = 1.0, 1.5, 2.0, 2.5) and wavelet decompositions.

The radiomic feature extraction process began with the use of 3D Slicer software (www.slicer.org/) to derive features from tumor image data. Using the Radiomics module, three-dimensional image features such as first-order, second-order, high-order, and filtering features were extracted (13, 14). First-order features describe global gray-level characteristics such as intensity distribution and energy. Second-order features quantify local gray-level dependencies, including autocorrelation, contrast, correlation, difference average, and difference entropy. High-order features reflect complex textural attributes, such as coarseness and heterogeneity. A total of 80 feature types were computed from each tumor ROI: first-order (18 features), 3D shape (16 features), 2D shape (14 features), gray-level co-occurrence matrix (24 features), gray-level run length matrix (16 features), gray-level size zone matrix (16 features), neighboring gray-tone difference matrix (5 features), and gray-level dependence matrix (14 features), yielding 1,223 total radiomic features. Preprocessing was performed to ensure feature quality and model compatibility. Initially, null values were imputed using the median. Features were then filtered using the intraclass correlation coefficient (ICC > 0.75) to retain those with high reproducibility. Finally, all retained features were normalized using Z-score transformation to a range of (−1, 1).

Effective radiomics features were selected from the training set using the Least Absolute Shrinkage and Selection Operator (LASSO) regression. Models were built using logistic regression, decision trees, support vector machines, and AdaBoost. Models were evaluated using accuracy, sensitivity, and specificity.

To develop a robust and efficient predictive model for the pathological grade of bladder cancer, diverse clinical factors were integrated. These included sex, age, height, weight, history of hypertension, urine pH, urinary tract infection, hematuria, urinary protein levels, triglycerides, BMI, diabetes, tumor location, tumor number, HDL-C, tumor size (maximum diameter), pathological grade, MS, past medical history, and family history. Given the data-intensive and complex nature of this study, categorical parameters were assigned numerical values to enhance computational efficiency and clarity in display. For instance, low-grade urothelial carcinoma was assigned a value of “0,” while other classifications were marked as “1.” For further details on the assignment of clinical factors, please refer to Table 1 .

Initially, four predictive models were developed to classify the pathological grade of bladder cancer using the pre-divided training set. Accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) were used to compare model performance. To assess the predictive power of these models, a 10-fold cross-validation procedure with 100 iterations was employed. In parallel, comparisons were made with a clinical feature-based prediction model to identify the model with superior predictive capability.

The software utilized in this study included 3D Slicer (version 4.10.2 for Windows 64-bit), RStudio (version 1.2.1335), and associated software packages. SPSS 22.0 (IBM) was used to analyze clinical variables, with continuous data expressed as mean ± standard deviation (x ± s). Independent samples t-tests were used to compare two groups of continuous variables, while chi-square tests were applied to categorical data. Logistic regression was used to identify independent risk factors. The effectiveness of radiological features in distinguishing between NMIBC and MIBC was evaluated using one-way analysis of variance. The LASSO regression model was implemented using the “glmnet” package, and ROC curves were plotted using the “pROC” package. Differences in AUC values between models were tested using the DeLong test, with statistical significance set at P < 0.05 (two-sided). Model performance was assessed using the concordance index (C-index), and decision curve analysis (DCA) was performed to evaluate clinical utility.

This study included 372 patients with bladder cancer. Of the participants, 301 were male and 71 were female, aged 26 to 95 years, with a mean age of 65.48 ± 12.39 years. The cohort comprised 182 individuals with low-grade urothelial carcinoma (LGUC) and 190 with high-grade urothelial carcinoma (HGUC). Table 2 presents the clinical characteristics of the high- and low-grade groups, as well as the overall cohort. Patients were randomly assigned to a training group (n = 259) and a validation group (n = 113) in a 7:3 ratio. Table 3 summarizes HGUC and LGUC data in both groups. Univariate analysis revealed statistically significant associations for age, urinary tract infection, hematuria, proteinuria, and tumor diameter (P < 0.05). Multivariate analysis identified age and proteinuria as independent risk factors in the clinical prediction model (P < 0.01).

A total of 699 radiomic features demonstrated strong inter- and intra-observer reproducibility, with ICC values of 0.75 or higher (mean ICC = 0.748; median = 0.807) (see Figure 3a ). One-way ANOVA identified 651 features with significant differences between HGUC and LGUC (P < 0.05), which were subsequently entered into a LASSO regression model for feature selection (see Figure 3b ). Using the minimum criterion from 10-fold cross-validation, 11 features were ultimately selected to build the predictive model (see Figures 3b, c ). Feature selection and modeling were performed using the “glmnet” and “pROC” packages in R.

Eleven radiomic features were finally selected for model construction: three first-order statistics (Entropy, Kurtosis, Skewness), five texture descriptors (GLCM_Contrast, GLCM_DifferenceVariance, GLRLM_LongRunEmphasis, GLDM_DependenceNonUniformity, NGTDM_Coarseness), and three shape indices (LongestAxisLength, SurfaceArea-to-Volume Ratio, Sphericity). Biologically, higher Entropy and Kurtosis indicate a disordered and peaked intensity distribution, often reflecting necrosis and marked cellular atypia characteristic of high-grade tumors. Texture metrics such as GLCM_Contrast and GLRLM_LongRunEmphasis quantify fine-scale grey-level fluctuations that mirror micro-architectural heterogeneity created by variable cell density, stromal content, and angiogenesis. Shape features capture macroscopic invasiveness; for example, an increased LongestAxisLength and reduced Sphericity signify elongation and irregular growth patterns commonly observed in aggressive lesions. Using these 11 features, four classifiers were evaluated—logistic regression, support vector machine (SVM), AdaBoost, and decision tree. Logistic regression showed the best overall clinical utility. In the training cohort, the AUCs were 0.858 (logistic regression), 0.872 (decision tree), 0.852 (SVM), and 0.789 (AdaBoost). For logistic regression, sensitivity, specificity, and accuracy in the training set were 75.40%, 82.71%, and 79.15%, respectively (95% CI: 0.812–0.903). In the validation set these metrics were 76.36%, 68.97%, and 72.57% (95% CI: 0.719–0.881). Thus, logistic regression maintained balanced, high diagnostic performance across both cohorts (AUC = 0.858 training; 0.800 validation; Figure 4 ). Although the AUC declined from 0.858 in the training set to 0.800 in the internal validation set, the overlapping 95% confidence intervals and well-aligned calibration curves suggest that significant overfitting did not occur. This modest reduction likely reflects normal sampling variability and underscores the importance of the forthcoming external validation study. Table 4 summarizes the diagnostic metrics for all models.

Univariate and multivariate analyses identified age and proteinuria as independent risk factors for bladder cancer pathological grade. A clinical prediction model incorporating these two variables was developed in the training group and compared with a radiomics-only model. The AUC values for the radiomics and clinical models were 0.858 and 0.650, respectively, with a statistically significant difference between them (P = 0.001). Figure 5 presents ROC curves for the clinical, radiomics, and combined models. The radiomics model outperformed the clinical model in predicting pathological grade. A comprehensive model integrating radiomics and clinical factors achieved improved performance (AUC = 0.864).

A radiomic nomogram was developed by integrating clinical factors (age and proteinuria) with the radiomics score (Rad-Score), as shown in Figure 6 . DCA indicated that the radiomics + clinical model demonstrated the highest net clinical benefit ( Figure 7 ). The calibration curve assessed the nomogram’s goodness of fit, where the 45° dashed line represents the ideal predictive model and the solid black line reflects the nomogram’s actual performance ( Figure 8 ). Closer proximity of the black line to the ideal diagonal indicates higher calibration accuracy.

To improve the clinical interpretability and predictive efficiency of the radiomics model, it was merged with the clinical model to generate a more intuitive and practical nomogram. Due to the complexity of incorporating 11 radiomic features individually, dimensionality was reduced by summarizing these features into a composite Rad-Score. The Rad-Score was calculated using the following formula:

Rad-scores: Log-sigma-1-0-mm-3D_glcm_Autocorrelation*1.03823

-Log-sigma-1-5-mm-3D_firstorder_Kurtosis*0.11197

+Log-sigma-2-0-mm-3D_glcm_JointAverage*0.12997

-Log-sigma-2-5-mm-3D_gldm_LowGrayLevelEmphasis*

-0.38671-Wavelet-LHH_glcm_Autocorrelation*1.72581

-Wavelet-LHH_glszm_SmallAreaHighGrayLevelEmphasis*0.88659

+Wavelet-LHL_gldm_LargeDependence*0.45270

+Wavelet-LLH_glcm_JointAverage*3.25086

-Original_shape_MajorAxisLength*0.05676

+Original_shape_MinorAxisLength*0.57564

+Original_shape_Maximum2DDiameterSlice*0.87998