JMIR Publications

ABSTRACT

Background:

Some researchers have explored the application of radiomics-based machine learning to detect preoperative muscle invasion, high-grade tumors, HER2 expression, and other risk factors for bladder cancer. However, systematic evidence proving its effectiveness remains lacking.

Objective:

The current study aimed to evaluate the performance of radiomics-based machine learning in preoperative risk stratification for bladder cancer patients. These findings could contribute to advancing the development or updating of intelligent risk assessment tools for bladder cancer.

Methods:

The Embase, Cochrane Library, PubMed, and Web of Science databases were systematically retrieved for publicly available studies on the effectiveness of radiomics-based machine learning in the preoperative risk stratification of bladder cancer up to October 17, 2025. Studies that only performed image segmentation or texture extraction without constructing a machine learning model were excluded from the primary study pool assessing risk of bias (ROB). The ROB in the included studies was evaluated using the Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST-AI). The overall quality of the studies was quantified employing the Radiomics Quality Scoring (RQS) tool. The certainty of the evidence was graded using the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) framework. Subgroup analyses were conducted according to the type of imaging source and modeling method.

Results:

This meta-analysis ultimately incorporated 57 studies with a total of 11,933 participants. These studies primarily utilized radiomics-based machine learning to identify muscle invasion (n = 34) and high-grade tumors (n = 16). Additionally, the methodology was employed to evaluate HER2-positive expression (n = 3), Ki-67 expression (n = 2), and lymph node (LN) staging (n = 2) preoperatively in bladder cancer. In the validation sets, the pooled area under the receiver operating characteristic curve (AUROC) for identifying muscle invasion was 0.900 (95% confidence interval [CI]: (0.863~0.937), 0.881 (95% CI: 0.857~0.906), and 0.840 (95% CI: 0.737~0.958) for computed tomography (CT)-, magnetic resonance imaging (MRI)-, and ultrasound (US)-based radiomics, respectively. The AUROC was 0.865 (95% CI: 0.840~0.891) and 0.915 (95% CI: 0.883~0.950) for models integrating clinical features with CT- or MRI-based radiomics, respectively. The pooled AUROC for diagnosing high-grade tumors was 0.818 (95% CI: 0.785~0.852) and 0.846 (95% CI: 0.663~1.000) for CT- and MRI-based radiomics, respectively. Furthermore, the AUROC was 0.883 (95% CI: 0.837~0.933) for MRI-based radiomics combined with clinical features. Among the included investigations, 16 studies (28%) implemented external validation.

Conclusions:

Despite the promising discriminatory performance of radiomics-based machine learning in detecting muscle invasion and classifying high-grade tumors in bladder cancer, the current evidence faces significant challenges, including methodological shortcomings and a high ROB, which currently preclude its readiness for clinical translation. Future studies will explore more robust radiomics methods for intelligent diagnosis.