MRI radiomics for the prediction of histologic grade in bladder cancer

Full Text

BACKGROUND

Bladder cancer (BC) is one of the most common malignant tumors of the urinary system. According to the GLOBOCAN report,¹ BC was the 9th most common cancer and the 13th by mortality rates in 2022. Russia is one of the top ten countries with the highest BC rates [1].

Many publications have covered the role of imaging techniques in diagnosing BC [2]. Currently, magnetic resonance imaging (MRI) with intravenous contrast enhancement is internationally recognized as the gold standard for non-invasive BC diagnosis [3]. According to Russian² and global [4] guidelines, the preferred diagnostic method for patients with suspected bladder neoplasms is cystoscopy and transurethral resection, followed by mandatory histopathological examination of tissues.

Clinical decisions are influenced by both the extent of the local tumor and tumor grade. According to global and Russian guidelines, a poorly differentiated (G3) tumor is considered high risk and requires additional or more radical treatment options.² Poorly differentiated BC is associated with lower overall and relapse-free survival rates [5].

Currently, BC is graded based on histopathological findings. The two most pressing issues in diagnosing BC are collecting all the necessary materials and conducting an accurate morphological examination. The diagnostic performance of transurethral resection depends on various factors, such as technical challenges, qualifications of medical personnel, high-energy tissue damage, tumor location, and resection technique [6, 7]. Efforts are being made to identify tumor grading markers in texture analysis that correlate well with histopathological findings. This is especially relevant in cases where intraoperative sampling does not provide diagnostic information.

The work aimed to evaluate the diagnostic value of texture analysis of magnetic resonance (MR) images in distinguishing between well- /moderately differentiated- (G1/2) and poorly differentiated (G3) bladder tumors; compare the results of 2D and 3D segmentation of MR images, compare various algorithms (LASSO, LASSO AutoML, Extra Trees, KNN, Random Forest, XGBoost, and LightXGB) in constructing models for predicting BC grade based on texture features, and select the most optimal algorithm based on main performance metrics.

METHODS

Study Design

This was a two-center retrospective study.

Study Setting

Data on patients with BC were provided by the Central Clinical Hospital with Inpatient Clinic of the Administrative Directorate of the President of the Russian Federation and the National Medical Research Center “Medical and Rehabilitation Center” of the Ministry of Health of Russia.

Eligibility Criteria

Inclusion criteria: available pelvic MR images of good quality; available histopathological findings with tumor grade documented.

Non-inclusion criteria: MR images and histopathological findings of low quality.

Exclusion criteria: absence of histopathological findings documented in the medical records.

Intervention

All patients underwent a standard preoperative pelvic MRI scan with intravenous contrast enhancement using a Siemens Magnetom Avanto 1.5 T (Siemens Healthineers, Germany), a Siemens Magnetom Aera 1.5 T (Siemens Healthineers, Germany), or a GE Signa HDxt 3.0 T (GE Healthcare, USA) scanners. Apparent diffusion coefficient (ADC) maps are automatically generated using diffusion-weighted images (DWI) with a high b-value (800/1000 s/mm).

Clinical and imaging data and texture features were analyzed. The analyzed clinical and imaging variables included age, sex, tumor grade, muscle layer involvement (based on histopathology reports), and VI-RADS (Vesical Imaging-Reporting and Data System) category.

LIFEx (v. 7.4.0, 7.5.5; www.lifexsoft.org) was used for calculations [8]. A radiologist with 4 years of imaging experience performed the image segmentation. In case of multiple lesions, the largest one was segmented. All images were preprocessed by resampling to a fixed voxel size of 1 mm × 1 mm × 1 mm. Three pulse sequences were manually segmented: T2-weighted images (T2-WI), DWI with a b-value of 800/1000 s/mm², and ADC, using a two-dimensional iterative region of interest (2D ROI) on a slice corresponding to the largest tumor size, and a three-dimensional (3D) ROI.

Sagittal T2-WIs were selected for segmentation. DWI and ADC sequences were acquired as axial slices. During segmentation, the ROI was transferred from one sequence to the other, followed by manual adjustment to correct any spatial misalignments. Quality control involved visual verification of the ROI alignment with tumor boundaries across all sequences following propagation and manual adjustment. The default texture parameter extraction settings were used for both the 2D ROI and the 3D ROI. After segmentation, 147 texture parameters were calculated for each pulse sequence. A total of 246 pulse sequences were analyzed.

Study Outcomes

The main study outcome was to identify radiomic parameters to differentiate between well-/moderately differentiated (G1/2) and poorly differentiated (G3) BC.

Additional study outcomes were to search for the optimal segmentation technique for MR images and the optimal algorithm for constructing a texture-based predictive model for BC grade.

Outcomes Registration. The outcomes were recorded through the analysis of MRI and histopathological reports, as well as texture analysis.

Statistical Analysis

Estimated sample size. The sample size was not predetermined.

Statistical methods. The results were statistically processed using Python version 3.12 (Python Software Foundation, www.python.org). Absolute values are presented as a median with an interquartile range (Me [Q1; Q3]), where Me is the median, and Q1 and Q3 are the first and third quartiles, respectively. Relative values are presented as a percentage per 100 examined patients. The two groups were compared using the Mann–Whitney test for absolute values and the chi-squared test for relative values. Differences were considered statistically significant at p ≤0.05.

The data from 95 patients was randomly divided into a training set and a test set (80:20). Several successive stages were used to select potential textural grade predictors. First, the texture analysis data was standardized. Then, univariate logistic regression was first used to identify the variables that significantly influenced the outcome. In the third stage, LASSO regression was used to select the most statistically significant features influencing the outcome collectively. The fourth stage involved using a machine learning algorithm to construct a model for predicting the presence or absence of the outcome. This algorithm allows for the construction of an easily interpretable, open model: multiple logistic regression (MLR). The following machine learning algorithms were also used to construct closed models: Extra Trees, KNN, Random Forest, XGBoost, and LightXGB. At the final stage, the constructed models were tested using an independent test set. The following performance metrics were used to evaluate the constructed models on the training and test sets: accuracy, sensitivity, specificity, and ROC-AUC, where ROC is receiver operating characteristic and AUC is area under the curve. The metrics were reported with 95% confidence intervals (CIs) estimated using the bootstrap method with 1000 resamples. For optimal sensitivity and specificity, a cutoff value was calculated to identify whether or not the study outcome was present. The study also employed a multi-model SHAP (SHapley Additive exPlanations) analysis to evaluate feature value.

RESULTS

Participants

Fig. 1 shows the sample selection process.

 

Fig. 1. Sample selection process. BC, bladder cancer; MRI, magnetic resonance imaging.

 

Sample Characteristics

The data of 98 patients with morphologically verified urothelial carcinoma (UBC) diagnoses were analyzed retrospectively. The study included data from 95 patients: 73 (76.8%) men and 22 (23.2%) women. In three cases, tumors could not be graded using histopathological findings. The mean age of the study participants was 68.4 [61.0; 75.7] years.

Based on the histopathological findings, well differentiated (G1) and moderately differentiated (G2) tumors, which are associated with low malignancy potential, were diagnosed in 56 patients (58.9%). Poorly differentiated (G3) tumors were reported in 39 patients (41.1%). Transurethral resection was performed no later than one month after MRI in 90 (94.7%) cases and within 1 to 3.5 months in 5 (5.3%) cases. Table 1 shows clinical and imaging characteristics of patients.

 

Table 1

Clinical and imaging characteristics of patients depending on tumor grade

Parameters		All patients examined n=95 (%)	G1/2 n=56 (%)	G3 n=39 (%)	p-value
Sex	male	73 (76.8)	40 (71.4)	31 (79.5)	>0.05
	female	22 (23.2)	16 (28.6)	8 (20.5)	>0.05
Age, years		68.4 [61.0; 75.7]	69.0 [61.0; 76.0]	69.0 [61.0; 76.0]	>0.05
Muscle invasion criterion n=81	MIBC	21 (25.9)	6 (12.2)	15 (46.9)	0.025
	NMIBC	60 (74.1)	43 (87.8)	17 (53.1)	0.025
VI-RADS n=94	VI-RADS (1/2/3)	70 (74.4)	47 (83.9)	23 (60.5)	0.03
	VI-RADS 4–5	24 (25.6)	9 (16.1)	15 (39.5)	0.03

Note. Table 1 and Table 2: MIBC, muscle-invasive bladder cancer; NMIBC, non-muscle-invasive bladder cancer; VI-RADS, Vesical Imaging-Reporting and Data System category; G1/2, well-/moderately differentiated tumor; G3, poorly differentiated tumor.

 

Primary Results

Clinical and imaging characteristics

There were no significant differences in age and sex characteristics. However, statistically significant differences (p ≤0.05) were obtained for some other parameters. In poorly differentiated (G3) tumors, muscle invasion was reported more often than in well differentiated (G1) and moderately differentiated (G2) tumors (p=0.025). Similar statistically significant differences were obtained for the VI-RADS categories. Among patients with poorly differentiated tumors (G3), 39.5% had a high potential for muscle invasion, according to MRI data (VI-RADS 4–5).

Performance of 2D texture analysis in predicting bladder cancer grade

The 2D texture analysis of MR images showed significant differences between the well- (G1)/moderately (G2) differentiated group and the poorly differentiated (G3) group. Statistically significant differences (p ≤0.05) were found in 65 features (16.2% of all analyzed features), including 20 features in the T2-WI sequence (30.8% of statistically significant features), 19 features in the DWI sequence (29.2%), and 65 features in the ADC sequence (40%).

The next step was to conduct a multivariate analysis after randomly dividing the patient data into a training set (n=76) and a test set (n=19) (Table 2). The groups demonstrated no statistically significant differences, indicating their comparability.

 

Table 2

Comparative characteristics of the training and test sets

Parameters		All patients examined n=95 (%)	Training set n=76 (%)	Test set n=19 (%)	p-value
Grade	G1/2	56 (58.9)	42 (55.3)	14 (73.4)	>0.05
	G3	39 (41.1)	34 (44.7)	5 (26.6)	>0.05
Muscle invasion criterion (n=81)	MIBC	60 (74.1)	46 (74.2)	14 (73.6)	>0.05
	NMIBC	21 (25.9)	16 (25.8)	5 (26.4)	>0.05
VI-RADS (n=94)	VI-RADS (1/2/3)	70 (74.4)	57 (75)	13 (68.4)	>0.05
	VI-RADS 4–5	24 (25.6)	18 (25)	6 (31.6)	>0.05

 

Table 3 shows the comparative performance metrics of various predictive models. The XGBoost model was found to be the most relevant based on performance metrics.

 

Table 3

Comparative performance metrics of predictive 3D models for bladder cancer grading

Models	ROC-AUC	Accuracy	Sensitivity	Specificity
LASSO	70.0% [44.1; 94.1]	57.9% [36.8; 73.7]	80.0% [50.0; 100.0]	50.0% [28.6; 71.4]
LASSO-AutoML	78.6% [60.3; 95.6]	52.6% [31.6; 73.7]	100.0% [100.0; 100.0]	35.7% [14.3; 58.3]
ET	82.1% [60.0; 100.0]	63.2% [42.1; 84.2]	80.0% [40.0; 100.0]	57.1% [35.3; 81.2]
KNN	75.7% [54.2; 94.3]	73.7% [57.9; 89.5]	40.0% [0.0; 80.0]	85.7% [69.2; 100.0]
RF	82.9% [61.1; 100.0]	57.9% [36.8; 78.9]	80.0% [40.0; 100.0]	50.0% [28.6; 75.0]
XGB	72.9% [39.7; 96.9]	78.9% [63.2; 94.7]	80.0% [40.0; 100.0]	78.6% [60.0; 100.0]
LightXGB	77.1% [56.7; 95.8]	68.4% [52.6; 84.2]	80.0% [40.0; 100.0]	64.3% [43.8; 85.7]

Note. Table 3 and Table 4: ET, Extra Trees; KNN, K-Nearest Neighbors; RF, Random Forest; XGB, eXtreme Gradient Boosting; LightXGB, light gradient boosting.

 

The XGBoost model included 40 texture features: 12 derived from T2-WI and 14 each from the DWI and ADC sequences. Notably, 80.0% of the texture variables were first-order features and shape descriptors, while 20.0% were second-order features (Supplement 1).

Three 2D texture features common to all models were identified: ADC_2D_INTENSITY_BASED_25thIntensityPercentile (the 25th percentile of the gray-level distribution within the region of interest); DWI_2D_MORPHOLOGICAL_MaxIntensityCoor (the maximum intensity coordinate); and ADC_2D_NGTDM_Coarseness (representing coarseness as the sum of differences between the gray level of a pixel/voxel and its neighbors, thus reflecting the spatial rate of change in intensity). All selection algorithms identified features derived from ADC and DWI sequences, highlighting the importance of these sequences in predicting tumor grade.

Next, SHAP analysis was performed to enhance model interpretability and evaluate the value of radiomic features for predicting the outcome (Supplement 2). Model coefficients describe the overall value of features, whereas SHAP values quantify the predictive value of individual features, confirming the consistency of their global ranking. The variables are listed in order of global importance, with the most important variable listed first. Each point on the horizontal axis represents a separate prediction within the set. The X-axis position shows whether a feature increases or decreases the prediction relative to the set average: the farther from zero, the greater the feature’s impact on a specific prediction. The scatter of the points along the horizontal axis shows the variability in the predictive value of the feature. The vertical color axis runs from blue to red and reflects the value of the feature, which ranges from low to high. For example, the ADC_2D_INTENSITY_BASED_MeanIntensity feature exhibits a cluster of red dots on the negative side of the SHAP values and blue dots on the positive side. This indicates that high values of this variable contribute negatively to the prediction, while low values have a strong positive impact. The ADC_2D_INTENSITY_BASED_25thIntensityPercentile feature exhibits the largest number of blue dots on the positive side, meaning that these features contributed to the final result at low values. Of the 23 most significant texture features considered, 18 were associated with ADC and DWI sequences (Supplement 2).

The study identified a radiomic profile (Rad-score) for each patient based on the generated predictive formula. The distribution of radiomic scores for each clinical case is presented as waterfall plots. A cut-off value of 0.45 was determined. If the predicted probability is <0.45, the tumor is G1/2; if ≥0.45, it is G3 (Supplement 3). The resulting model demonstrated good performance (70%–80%), with a ROC-AUC of 72.9% on the test set. A confusion matrix was constructed (Supplement 4). The plots demonstrate that one false result was recorded in the training set (one green bar is located below the 0.45 cut-off). The model incorrectly predicted the outcome in 4 cases in the test set.

Performance of 3D texture analysis in predicting bladder cancer grade

The 3D texture analysis of MR images showed significant differences between the low (G1)/intermediate (G2)-grade group and the high (G3)- grade group. Statistically significant differences (p ≤0.05) were found in 97 features (24.3% of all analyzed features), including 36 features in the T2-WI sequence (37.1% of statistically significant features), 30 features in the DWI sequence (30.9%), and 31 features in the ADC sequence (32.0%).

In summary, we have preliminarily concluded that MRI texture analysis has a high informative value in predicting tumor grade.

The next step was to conduct a multivariate analysis after randomly dividing the patient data into a training set (n=76) and a test set (n=19) (Table 2). The groups demonstrated no statistically significant differences, indicating their comparability.

Table 4 shows the comparative performance metrics of various predictive models. Based on performance metrics, we concluded that the XGBoost model was the most relevant.

 

Table 4

Comparative performance metrics of predictive 3D models for bladder cancer grading

Models	ROC-AUC	Accuracy	Sensitivity	Specificity
LASSO	72.9% [51.4; 93.3]	57.9% [36.8; 78.9]	80.0% [40.0; 100.0]	50.0% [27.3; 73.3
LASSO-AutoML	81.4% [62.9; 97.4]	63.2% [47.4; 78.9]	100.0% [100.0; 100.0]	50.0% [28.6; 72.7]
ET	75.7% [52.9; 95.8]	52.6% [36.8; 73.7]	80.0% [40.0; 100.0]	42.9% [21.4; 66.7]
KNN	90.0% [75.6; 100.0]	73.7% [57.9; 89.5]	20.0% [0.0; 57.1]	92.9% [78.6; 100.0]
RF	85.7% [67.6; 100.0]	68.4% [52.6; 84.2]	80.0% [40.0; 100.0]	64.3% [43.8; 85.7]
XGB	82.9% [59.4; 100.0]	78.9% [63.2; 94.7]	80.0% [40.0; 100.0]	78.6% [58.8; 94.1]
LightXGB	77.1% [56.7; 95.8]	68.4% [52.6; 84.2]	80.0% [40.0; 100.0]	64.3% [43.8; 85.7]

 

The XGBoost model incorporated 23 texture features, including 10 derived from the T2-WI sequence, 2 from DWI, and 11 from ADC maps. Notably, 69.6% of the texture variables were first-order features and shape descriptors, while 30.4% were second-order features (Supplement 5).

The variable common to all models was T2_3D_INTENSITY_HISTOGRAM_AreaUnderCurveIVH (Area under the curve of cumulative intensity volume histogram), which describes the relationship between the discretized gray level and the fraction of the region of interest containing that gray level.

A cut-off value of 0.45 was determined. If the predicted probability is <0.45, the tumor is G1/2; if ≥0.45, it is G3. The resulting model performs very well (80%–90%), with a ROC-AUC of 82.9% on the test set. The study identified a radiomic profile (Rad-score) for each patient based on the generated predictive formula (Supplement 6). A confusion matrix was constructed (Supplement 7). At the cut-off point of 0.45, the number of correctly predicted outcomes by the model reached 98.9% and 78.9% in the training and test sets, respectively.

DISCUSSION

Summary of Primary Results

Based on the analysis of the main performance metrics, the XGBoost model showed the best discriminative performance. XGBoost is an ensemble regression model. Implemented using gradient boosting technology, it is a popular method for constructing predictive models, including those used in texture analysis [9]. Gradient Boosting is a machine learning technique that builds a predictive model in the form of an ensemble of decision trees. Ensemble training is performed sequentially rather than in parallel [10]. First, the initial ensemble model (tree 1) is trained on all available data. Then, this model is tested on a test dataset to identify errors in its performance. The identified errors in tree 1 are then used to train the next model in the ensemble, tree 2. Tree 2 is trained to compensate for the errors of tree 1. Errors from the previous model are used to train the next one. Therefore, each subsequent model added to the ensemble is designed to compensate for the errors arising in the previous stages.

Interpretation

Our study compared 2D and 3D XGBoost predictive models based on texture analysis results and found that both performed well. However, the 3D radiomic model had a slight advantage. Although accuracy, sensitivity, and specificity were similar, the 3D model’s ROC-AUC was 82.9%, compared to 72.9% for the 2D model. In addition, the 3D model demonstrated fewer texture variables and a simpler predictive formula, both of which are convenient for clinical practice. However, the 3D model has one disadvantage: the image segmentation process is more time-consuming and labor-intensive.

We demonstrated that the ratio of texture predictors differs between the 2D and 3D models. In the 2D model, second-order texture features (79.5%) dominated, reflecting spatial relationships between pixels within a single slice. In contrast, in the 3D model, most texture predictors (69.6%) consisted of first-order features, which describe the global distribution of voxel intensities throughout the entire tumor volume. In our opinion, this difference suggests that 2D segmentation is better for analyzing complex spatial patterns within a single slice. However, 3D segmentation is more sensitive to the overall histogram characteristics of the entire lesion volume. It can also be concluded that both segmentation methods provide a more comprehensive, multidimensional characterization of texture patterns associated with tumor grades. Notably, similar conflicting findings have been reported in several previous papers [11–14].

We did not include ADC cutoff values in the analysis because they are not standardized and vary significantly depending on the equipment used [15, 16]. We also did not incorporate clinical or imaging parameters into the predictive models.

Following the approach of other authors [17, 18], we limited our texture analysis to a multiparametric approach. We did not aim to compare the predictive potential of models based on individual pulse sequences. Our findings are consistent with previous results regarding the significance of texture features from ADC sequences in distinguishing between poorly and well-differentiated tumors [12, 14, 18], as well as with data on discriminative characteristics [17, 18].

Although texture analysis does not eliminate the need for histopathological examination, it can support preoperative patient stratification when invasive diagnostic options are limited. This elevates the diagnostic algorithm to a new level. Texture analysis of tumor grades has the potential to facilitate the transition to more personalized patient management strategies.

Study Limitations

A limitation of this study is its retrospective design, which could potentially introduce selection bias.

The MR images included in the study were obtained at different sites using various scanners, which does not rule out the possibility of technical interference.

The relatively small sample size, especially in the test set, may have affected the stability of the model metric estimates. We applied only voxel size normalization for study standardization and did not utilize various filters or gray-level scale constraints. The generalizability of our results may be limited due to the lack of external validation using an independent patient cohort from another site. The texture analysis was limited to the tumor itself. However, analyzing the structure of the surrounding tissue could be of additional interest. Image segmentation by a single radiologist precludes the assessment of inter-rater reliability.

CONCLUSION

Texture analysis of MR images can be used to grade BC, and the XGBoost algorithm outperforms other algorithms in predictive models. The difference in the texture predictor category ratio between 2D and 3D predictive models is of significant methodological importance because of their high accuracy and comparability. This finding demonstrates that the two segmentation approaches are complementary and the same diagnostic result can be achieved by analyzing different aspects of tumor texture heterogeneity.

ADDITIONAL INFORMATION

Supplement 1. Selected 2D-radiomics features and their coefficients. The values of the coefficients are used to make the prediction formula.

doi: 10.17816/clinpract696013-4415221

Supplement 2. SHAP-analysis. Graphical representation of the 23 most important 2D-texture features, whose contribution is reduced from top to bottom.

doi: 10.17816/clinpract696013-4415223

Supplement 3. 2D radiomic profile of patients in the training (A) and test (B) sets.

doi: 10.17816/clinpract696013-4415224

Supplement 4. Confusion matrix for the 2D radiomic XGBoost predictive model on training and test sets. The actual and predicted subtype classifications are shown on the vertical and horizontal axes, respectively.

doi: 10.17816/clinpract696013-4415225

Supplement 5. XGBoost-based 3D predictor coefficients. These values are used for the predictive model.

doi: 10.17816/clinpract696013-4415226

Supplement 6. 3D radiomic profile of patients in the training (A) and test (B) sets.

doi: 10.17816/clinpract696013-4415227

Supplement 7. Confusion matrix for the 3D radiomic XGBoost predictive model on training and test sets. The actual and predicted subtype classifications are shown on the vertical and horizontal axes, respectively.

doi: 10.17816/clinpract696013-4415229

Author contributions: А.А. Kovalenko, general concept, search and analytical work, writing the text of the article; V.E. Sinitsyn, general concept, editing the text; V.S. Petrovichev, search and analytical work, editing the text. Thereby, all authors provided approval of the version to be published and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: The research protocol was approved by the Ethics Committee at the National Medical Research Centre “Treatment and Rehabilitation Centre” (extract from Protocol No. 086 dated 16.09.2025). All the study participants voluntarily signed an informed consent form before MRI.

Funding source: The study had no sponsorship.

Disclosure of interests: The authors declare that they have no competing interests.

Statement of originality: The authors did not utilize previously published information (text, illustrations, data) in conducting the research and creating this paper..

Data availability statement: The editorial policy regarding data sharing does not apply to this work, data can be published as open access.

Generative AI: Generative AI technologies were not used for this article creation.

 

¹ International Agency for Research on Cancer. Cancer today. Data visualization tools for exploring the global cancer burden in 2022. Available at: http://gco.iarc.fr/today/home Accessed on February 15, 2026.

² Clinical Guidelines. С67. Malignant Neoplasm of Bladder. The Russian Association of Oncological Urology et al., 2023. Available at: https://oncology-association.ru/wp-content/uploads/2023/11/rak-mochevogo-puzyrya_23.pdf Accessed on February 15, 2026.

About the authors

Anastasia A. Kovalenko

Central Clinical Hospital, Moscow

Author for correspondence.
Email: nastua_kovalenko@mail.ru
ORCID iD: 0000-0001-8276-3594
SPIN-code: 6158-0090
Russian Federation, Moscow

Valentin E. Sinitsyn

Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies; Lomonosov Moscow State University

Email: vsini@mail.ru
ORCID iD: 0000-0002-5649-2193
SPIN-code: 8449-6590

MD, PhD, Professor

Russian Federation, Moscow; Moscow

Victor S. Petrovichev

National Medical Research Centre «Treatment and Rehabilitation Centre»

Email: petrovi4ev@gmail.com
ORCID iD: 0000-0002-8391-2771
SPIN-code: 7730-7420

MD, PhD

Russian Federation, Moscow

References

Supplementary files