MRI-derived features reveal recurrence heterogeneity in VI-RADS 2 non-muscle-invasive bladder cancer

To characterize recurrence heterogeneity among patients with VI-RADS 2 non-muscle-invasive bladder cancer (NMIBC) and to develop and internally validate a recurrence-free survival (RFS) prediction model incorporating quantitative MRI features.

This retrospective study included 342 patients with pathologically confirmed VI-RADS 2 NMIBC, who were randomly divided into training and internal validation cohorts. Clinical and quantitative MRI variables were evaluated using Cox regression, least absolute shrinkage and selection operator regression, Boruta, and recursive feature elimination. Six survival prediction algorithms were developed and compared. The selected model was further compared with a clinical Cox model and the European Association of Urology (EAU) risk stratification system using the concordance index (C-index), time-dependent area under the curve (AUC), calibration, Brier score, and decision curve analysis.

Among 342 patients, 111 (32.5%) experienced recurrence. Six predictors were retained: tumor number, maximum tumor vertical distance, maximum tumor diameter, maximum tumor-base contact length, mean apparent diffusion coefficient (ADC) at the tumor base, and minimum ADC within the tumor. The random survival forest (RSF) model achieved C-index values of 0.832 and 0.757 in the training and validation cohorts, respectively. In the validation cohort, the corresponding C-index values were 0.417 for the clinical Cox model and 0.534 for the EAU risk stratification system. The RSF model also yielded higher time-dependent AUCs than both reference models and greater net benefit at the 60-month horizon. RSF-derived risk groups showed significantly different RFS in both cohorts; in the validation cohort, the hazard ratio for the high-risk versus low-risk group was 5.30 (95% confidence interval, 2.39-11.76).

Within this single-center cohort, recurrence risk varied substantially among patients with VI-RADS 2 NMIBC. An RSF model incorporating quantitative MRI features and tumor number showed better internal discrimination than the clinical Cox model and the EAU risk stratification system and separated patients into distinct risk groups. These findings suggest that quantitative MRI features may complement conventional clinical assessment in this imaging-defined subgroup. Prospective multicenter external validation with longer follow-up is required before the model can be considered for clinical use.