Biparametric MRI with deep learning for the preoperative prediction of muscle invasion in patients with bladder cancer: a multicenter study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, The Tenth Affiliated Hospital of Southern Medical University (Dongguan People's Hospital), Dongguan, China.
- Department of Radiology, Daojiao Hospital of Dongguan, Dongguan, China.
- Medical AI Lab, School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, Shenzhen, China.
- Department of Radiology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China.
- Imaging Department, Sun Yat-Sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangzhou, China.
- Medical AI Lab, School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, Shenzhen, China. [email protected].
- Department of Radiology, The Tenth Affiliated Hospital of Southern Medical University (Dongguan People's Hospital), Dongguan, China. [email protected].
- Department of Radiology, The Tenth Affiliated Hospital of Southern Medical University (Dongguan People's Hospital), Dongguan, China. [email protected].
Abstract
This study aimed to develop and validate a deep learning model that integrates T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) for the noninvasive prediction of muscle invasion in bladder cancer. A total of 308 bladder tumors from 254 patients were retrospectively collected from three institutions. Of these, 177 tumors from Center 1 were partitioned via stratified sampling into a training set (n = 128) and an internal validation set (n = 49). The remaining two centers contributed 59 and 72 tumors, respectively, which served as external validation cohorts. A dual-branch 3D deep convolutional neural network (Joint model) based on biparametric MRI (bp-MRI) was developed using the training data and compared with a deep learning model based on T2WI (T2WI model) and VI-RADS assessments by radiologists. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The Joint model achieved mean AUCs of 0.979 ± 0.007 in the internal validation cohort, and 0.907 ± 0.013 and 0.916 ± 0.021 in the external validation cohorts. It outperformed the T2WI model in all cohorts and achieved higher AUCs than radiologists across all validation cohorts. Notably, for equivocal VI-RADS 2-3 lesions, the Joint model yielded higher accuracy than the T2WI model in external validation. Compared with the T2WI model and VI-RADS assessment, the Joint model demonstrated better diagnostic performance and generalizability for predicting muscle invasion in patients with bladder cancer. Question Reliable noninvasive tools for preoperative prediction of muscle invasion in bladder cancer are lacking, as TURBT-based staging carries procedural risks and frequent understaging. Findings A biparametric MRI deep learning model achieved AUCs of 0.907-0.979 across internal and external validation cohorts, outperforming T2WI-only models and VI-RADS. Critical relevance statement Contrast-free biparametric MRI combined with deep learning enables accurate, generalizable preoperative prediction of muscle invasion, offering particular value for equivocal VI-RADS 2-3 lesions and supporting noninvasive clinical decision-making.