[MRI-based deep learning model for preoperative prediction of urothelial carcinoma with variant histology of bladder:...

Objectives: To develop a deep learning model based on magnetic resonance imaging (MRI) for the preoperative prediction of urothelial carcinoma with variant histology (VUC), and to evaluate its predictive performance and prognostic stratification value. Methods: This is a retrospective cohort study. Data from 1 221 patients with bladder cancer across seven centers of China between June 2013 and May 2025 were retrospectively analyzed. There were 1 032 male cases and 189 female cases, with an age (M(IQR)) of 68.0 (15.0) years (range: 19 to 96 years). Data from Jiangsu Provincial People's Hospital (n=959) were randomly assigned to a training set (n=767) and a validation set (n=192) in an 8∶2 ratio using stratified random sampling method, while data from the other six centers (n=262) served as an external test set. Patients were categorized into pure urothelial carcinoma (PUC) group and VUC group. A deep learning prediction model for VUC (DeepVUC) was developed using a 3D ResNet50-based framework, integrating MRI T2WI and diffusion-weighted imaging (DWI) through a self-attention mechanism. The comparison of data between groups was conducted using Mann-Whitney U, χ² test or Fisher's exact probability method. Pairwise comparisons among multiple groups were corrected using the Bonferroni method. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), and the prognostic value was assessed through Kaplan-Meier analysis. Results: Among the 1 221 patients with bladder cancer, 1 078 cases were PUC and 143 cases were VUC, with a proportion of VUC at 11.7%. Comparative analysis revealed that relative to the PUC group, the proportions of lymph node metastasis, high-grade urothelial carcinoma, carcinoma in situ, and lymphovascular invasion were significantly higher in the VUC group (all P<0.01). Bonferroni pairwise comparison showed that the proportion of VUC in pT3 stage patients was significantly higher than that in pT2 stage (χ²=14.30, P<0.01), pT1 stage (χ²=56.76, P<0.01) and pTa stage (χ²=174.02, P<0.01). The VUC group also exhibited significantly higher vesical imaging-reporting and data system(VI-RADS) scores (χ²=141.48, P<0.01). Specifically, the percentage of VUC in patients with VI-RADS scores of 1 to 5 was 3.5% (4/113),4.2% (22/524), 9.1% (25/275), 25.0% (37/148) and 34.2% (55/161), respectively. Squamous differentiation was the most common variant, accounting for 51.7% (74/143) of all VUC patients, with the highest frequency observed in patients with a VI-RADS score of 5 (69.1%, 38/55). The DeepVUC model achieved AUCs of 0.972 (95%CI: 0.955 to 0.989), 0.940 (95%CI: 0.902 to 0.977), and 0.925 (95%CI: 0.889 to 0.961) in the training, validation, and external test sets, respectively. Accuracies were 95.2%, 90.1%, and 84.4%; sensitivities were 92.1%, 90.9%, and 84.4%; and specificities were 95.6%, 90.0%, and 84.3%. Prognostic analysis showed that the DeepVUC-defined high-risk group had significantly worse overall survival compared to the low-risk group in the total cohort (HR=4.51, 95%CI: 3.24 to 6.26, P<0.01), the PUC group (HR=2.13, 95%CI: 1.42 to 3.18, P<0.01), and the VUC group (HR=3.17, 95%CI: 1.75 to 5.76, P<0.01). Conclusions: The DeepVUC model enables non-invasive preoperative prediction of VUC in bladder cancer and demonstrates significant value in prognostic stratification. As a non-invasive imaging biomarker, this model provides a valuable reference for the development of individualized treatment strategies for bladder cancer patients.