The distinction matters because treatment pathways diverge sharply at this boundary. Non-muscle-invasive tumors are typically managed with transurethral resection followed by intravesical therapy and surveillance, whereas muscle-invasive disease ordinarily warrants radical cystectomy or trimodality bladder-preservation therapy with neoadjuvant chemotherapy. Misclassification in either direction carries serious consequences: undertreatment of invasive disease, or unnecessary radical surgery with substantial impacts on survival and quality of life. Current staging relies on cystoscopy, cross-sectional imaging and transurethral resection, but even deep biopsy is limited by sampling error, and a meaningful proportion of high-risk patients are found to have more advanced disease at radical cystectomy.

To standardize magnetic resonance imaging of the bladder, the Vesical Imaging-Reporting and Data System, or VI-RADS, was introduced in 2018. Meta-analyses have shown the system to be reproducible, with pooled sensitivities of roughly 88 to 90 percent and specificities of 80 to 85 percent for predicting muscle invasion at a cut-off score of 3 or higher. Yet scores of exactly 3 remain a recognized gray zone, and performance varies with reader experience and tumor location. Radiomics offers a complementary route: by extracting high-dimensional quantitative descriptors of intensity, texture and shape from medical images, it can capture intratumoral heterogeneity that is not discernible to the human eye.

The research team, led by investigators at the PLA Rocket Force Characteristic Medical Center in Beijing and Qianjiang Central Hospital in Hubei, China, analyzed consecutive patients who underwent preoperative 3.0-Tesla multiparametric MRI between January 2019 and December 2024. Of 412 patients initially identified, 62 were excluded for reasons including severe imaging artifacts, non-urothelial histology, prior treatment and incomplete data, leaving 350 for analysis. Rather than splitting patients randomly, the investigators divided the cohort chronologically: 245 patients imaged through June 2023 formed the training cohort, and 105 patients imaged from July 2023 to December 2024 served as a temporal validation cohort, a design intended to better mimic real-world deployment of the model.

The radiomics pipeline drew on T2-weighted imaging, high-b-value diffusion-weighted imaging and apparent diffusion coefficient maps. Radiologists manually delineated whole-tumor volumes of interest on each sequence, and more than 4,200 candidate features per patient were extracted using the open-source PyRadiomics software. Rigorous quality control followed: features had to show intra- and inter-reader reproducibility with intraclass correlation coefficients of at least 0.80, a threshold that removed roughly a third of candidates. After univariate screening and pruning of highly collinear features, 64 features entered least-absolute-shrinkage-and-selection-operator regression, which retained eleven features with non-zero coefficients: four from T2-weighted imaging, four from ADC maps and three from diffusion-weighted imaging. These were combined into a continuous Rad-score for each patient.

Muscle invasion was confirmed in 78 of 245 training patients, or 31.8 percent, and 33 of 105 validation patients, or 31.4 percent. The reference standard was detrusor muscle infiltration on the highest-quality definitive pathology available, with radical cystectomy specimens taking precedence over transurethral resection where both existed. Among the 107 patients with paired assessments, classification was discordant in 17.8 percent of cases, with 17 patients upstaged from non-muscle-invasive disease at resection to muscle-invasive disease at cystectomy. On multivariable analysis, five predictors remained independently associated with invasion: the Rad-score, the VI-RADS score, tumor size, the presence of hydronephrosis, and the absence of a visible tumor stalk on MRI.

The combined nomogram delivered an area under the curve of 0.897 in temporal validation, with a sensitivity of 84.8 percent and a specificity of 86.1 percent at a conventional probability threshold of 0.50. Calibration held up well: the validation calibration slope was 0.91 and the intercept 0.06, and the calibration curves stayed close to the ideal line, with Brier scores of 0.092 in training and 0.105 in validation. Decision curve analysis showed greater net benefit than treat-all, treat-none and single-model strategies across a broad range of threshold probabilities, though the authors stress that the model’s probabilities should inform multidisciplinary discussion rather than independently dictate treatment.

Several sensitivity analyses tested the robustness of the findings. Restricting the analysis to the 252 patients whose staging was confirmed by radical cystectomy left performance essentially unchanged, with validation areas under the curve of 0.892 for the combined model. Adjusting radiomics features for tumor volume, adding scanner type as a covariate, and excluding 14 resection specimens with inadequate detrusor-muscle sampling all produced nearly identical results. Variance inflation factors for all combined-model predictors stayed below 2.0, arguing against severe multicollinearity, and reduced models omitting individual predictors retained validation areas under the curve between 0.882 and 0.886, indicating that no single variable carried the entire signal.

The most clinically intriguing result came from the 59 patients whose research VI-RADS score was exactly 3, the diagnostically uncertain gray zone. Because the VI-RADS score is constant within this subgroup, it cannot by itself discriminate between invasion states. The combined nomogram achieved an area under the curve of 0.871 in the 39-patient training subgroup and 0.835 in the 20-patient validation subgroup, with a pooled descriptive estimate of 0.864. At a probability threshold of 0.50, pooled sensitivity, specificity and accuracy were 81.8, 81.1 and 81.4 percent respectively. The authors emphasize that these subgroup findings are exploratory and hypothesis-generating, given that the validation subgroup contained only seven muscle-invasive cases and correspondingly wide confidence intervals.

The study’s limitations are laid out in detail. All data came from a single tertiary center, and temporal validation within the same institution is not equivalent to external validation. The clinical-radiologic model showed marked deterioration between training and validation, a reminder that prediction models can be fragile when transported forward in time. The pathological reference standard mixed resection and cystectomy specimens, and because the decision to perform cystectomy relates to disease severity, the cystectomy-restricted analysis remains vulnerable to verification bias. Dynamic contrast-enhanced MRI was used for VI-RADS scoring but excluded from the radiomics signature to avoid variability from contrast timing, potentially omitting complementary perfusion information, and manual segmentation remains time-consuming. The authors conclude that prospective multicenter validation and possible recalibration are required before the nomogram can be implemented clinically, positioning the work as a strong methodological step toward preoperative risk stratification rather than a ready-made clinical tool.

Subject of Research: Multiparametric MRI radiomics combined with VI-RADS and clinical features for preoperative prediction of muscle-invasive bladder cancer

Article Title: Multiparametric MRI radiomics combined with clinical-radiologic features improves preoperative prediction of muscle invasion in bladder cancer

Article References: Qu, L., Qu, Z., Ye, X., Yuan, G., & Wang, L. (2026). Multiparametric MRI radiomics combined with clinical-radiologic features improves preoperative prediction of muscle invasion in bladder cancer. Frontiers in Oncology, 16, Article 1949830. https://doi.org/10.3389/fonc.2026.1949830

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1949830

Keywords: bladder cancer, multiparametric MRI, radiomics, VI-RADS, muscle invasion, nomogram, temporal validation, machine learning, urothelial carcinoma, prediction model, VI-RADS 3 gray zone, decision curve analysis