This study followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines (Checklist 1). The review was prospectively registered with PROSPERO (International Prospective Register of Systematic Reviews) under the title “Assessment of Preoperative Risk Stratification for Bladder Cancer Using Machine Learning Based on Radiomics: A Systematic Review and Meta-Analysis” (CRD42024561649).

The Cochrane Library (CENTRAL), Embase, PubMed, and Web of Science were retrieved up to June 17, 2024. This investigation did not perform simultaneous searches across multiple databases on a single platform. Searches were not extended to dedicated conference abstract databases or web sources beyond the primary databases. The authors of unpublished conference abstracts were not contacted to obtain full study details. To mitigate the risk of omitting relevant studies, the reference lists of eligible articles and pertinent review papers were manually scrutinized. No search filters were applied. The search strategy integrated both controlled vocabulary, such as Medical Subject Headings and Emtree terms, and free-text keywords to optimize sensitivity. Subject headings included “Urinary Bladder Neoplasms” and “machine learning.” Search strategies were tailored to the specific syntax of each database and combined using Boolean operators. No restrictions were applied concerning publication date or geographical location. Before the final data analysis, an updated search was performed in all specified databases on October 17, 2025 (Multimedia Appendix 1).

EndNote was used to import the retrieved articles. After eliminating duplicate records, the remaining articles were reviewed by titles and abstracts. Exclusions were categorized as follows: meta-analyses or reviews, replies or letters, case reports, animal experiments, registry or clinical trial protocols, non-English articles, and preprints. Initially relevant articles were screened by full text to determine eligible studies.

Before data extraction, a spreadsheet was created. The collected information encompassed DOI, title, publication year, first author, country, study type, purpose of the task, patient source, image source, recording of a complete image acquisition protocol, number of researchers involved, whether preliminary experiments tested different imaging parameters, radiomic region of interest (ROI) segmentation software, total number of outcome events, whether test-retest experiments were performed, total number of cases, training set size, number of outcome events in the training set, validation set generation method, validation set size, number of outcome events in the validation set, variable selection method, model types used, overfitting assessment method, public availability of data and code, mean age, sex, specimen source, AUROC (95% CI), number of true negatives, number of true positives, specificity, sensitivity, precision, accuracy, and F₁-score.

Results from each dataset within a study were included only once. When multiple studies published by the same author over different years were suspected of having overlapping datasets, only the study with the larger sample size was included. When multiple models were present, the model demonstrating optimal performance in the validation set was selected for inclusion. If a single dataset contained multiple validation sets, all were incorporated into the analysis.

Two researchers (ZH and YL) screened the articles and extracted the data separately. Their results were cross-checked. Any discrepancies were addressed with the help of a third researcher (BX).

The Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST-AI) provides a framework to critically appraise the risk of bias (ROB) and applicability of ML-based multivariable prediction models [23,24]. The assessment for ROB comprises four domains: participants and data sources, predictors, outcome, and analysis (1=strongly disagree, 2=somewhat disagree, 3=I don’t know, 4=somewhat agree, 5=strongly agree). A domain was judged as high risk if any signaling question was rated “no/probably no,” and as low risk only if all questions were answered “yes/probably yes.” An “I don’t know” response yielded an unclear risk designation.

Furthermore, the quality of the eligible investigations was appraised using the Radiomics Quality Score (RQS), which ranges from –8 to 36 [25]. This RQS primarily considers image protocol quality, multiple image segmentation methods, study of modalities, image acquisition time, feature reduction, and model construction using radiomic and nonradiomic features (prognosis and molecular subtyping). It also considers the detection and discussion of radiomic and biological correlations, cutoff value analysis, calibration statistics, discrimination statistics, validation, prospective studies registered in trial databases, comparison with the “gold standard,” cost-effectiveness analysis, potential clinical utility, and open science and data.

Two investigators independently evaluated the ROB and methodological quality and then performed a cross-check of their assessments. Any discrepancies were adjudicated with the help of a third researcher (BX).

A meta-analysis was performed on AUROC, a metric for the accuracy of ML models. The analysis required the SE or 95% CI of the AUROC. However, some of the included studies lacked a 95% CI and SE. In this scenario, SE was estimated by referring to the study by Debray et al [26]. A random-effects model was used for the meta-analysis to account for potential heterogeneity among the primary studies due to variations in model parameter tuning and predictor selection. Furthermore, the Hartung-Knapp-Sidik-Jonkman method was applied [27].

Furthermore, a meta-analysis was performed using a bivariate mixed effects model to pool sensitivity and specificity. This analysis was based on diagnostic contingency tables. Nevertheless, these tables were not directly provided in some studies. Then, the required values were calculated via reported sensitivity, specificity, precision, and sample sizes. Subsequently, subgroup analyses were conducted based on the dataset (training and validation sets), image source, and model type. A meta-analysis of AUROC values was conducted using packages “meta” and “metafor” in R (v4.5.2; R Foundation for Statistical Computing). A bivariate mixed effects model was implemented using the “midas” package in Stata (v15.0; StataCorp LLC).

Parameter definitions: c denotes the c-statistic; n represents the number of observed events; and m corresponds to the total sample size n∗=m∗=m+n2−1.

Finally, subgroup analyses were conducted within each task category based on case number, ML type (logistic regression [LR] vs other ML), and modeling variables (radiomics features alone vs radiomics features+clinical characteristics).

The certainty of evidence for the AUROC estimates derived from various subgroup analyses was evaluated using the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) framework. Two investigators independently performed the evidence grading and a cross-verification of their assessments. Any discrepancies were adjudicated by a third investigator (BX).

The database search yielded 8272 articles. After removing 2234 duplicates, 6038 articles remained and were reviewed based on their titles and abstracts. This process resulted in the exclusion of 5971 ineligible records. Overall, 67 articles remained and were reviewed by full texts, leading to the removal of 10 for irrelevance, including 4 non–peer-reviewed conference papers, one study lacking radiomics analysis, one focused on differentiating pure urothelial carcinoma from urothelial carcinoma with squamous differentiation, one without radiomics, one enrolling participants without a pathological diagnosis, one differentiating T2 from T3 bladder cancer stages, and one predicting programmed death-ligand 1 expression. Ultimately, the meta-analysis incorporated 57 articles [28-84] (Figure 1).

Fifty-seven [28-84] eligible studies, published between 2017 and 2025, encompassed data on 11,933 individuals with bladder cancer. Most were case-control studies, including 40 [28,33-36,38-41,43-46,48,49,51,53-57,61-76,79,82,84] single-center and 17 [29-32,37,42,47,50,52,58-60,77,78,80,81,83] multicenter studies. Thirty-four studies primarily investigated muscle invasion; 20 [30,32,33,41-43,46,51,52,54,57-60,63,64,69,71,78,83], 13[28,29,31,34,45,48,50,62,75,79-82], and one [65] used MRI-, CT-, and ultrasound-based radiomics, respectively. Sixteen studies [36,38-40,44,47,49,53,56,61,66,67,70,72,73,84] focused on discriminating high-grade tumors. Three studies [55,74,77] examined HER2-positive expression, 2 [35,76] examined Ki-67 expression, and 2 [37,68] examined LN staging in bladder cancer. ITK-SNAP (University of Pennsylvania) was the most frequently used software for ROI segmentation. Other software platforms included MATLAB (R2012b; Matrix Laboratory; MathWorks), the open-source Medical Imaging Interaction Toolkit, and a computer-assisted visualization and analysis software system. Fifty-three studies explicitly described the generation of a validation; among these, 37 [28,34-37,39-41,43-46,48,49,51,53-57,61-66,68-72,74-76,79,82,84] used internal validation techniques such as random sampling or cross-validation and 16 [29,30,31,32,42,47,50,52,58-60,77,78,80,81,83] studies used external validation. A total of 10 distinct model types were evaluated (Table 1).

Subgroup analyses of AUROC for the diagnosis of muscle invasion were performed based on different imaging modalities and model types. Among CT-based radiomics models, deep learning (DL) was the predominant approach. In the validation sets, the pooled AUROC of DL was 0.903 (95% CI 0.826-0.988; grade: weak), with a sensitivity of 0.88 (95% CI 0.77-0.94) and a specificity of 0.83 (95% CI 0.71-0.91). Among MRI-based radiomics models, DL was the most common modeling approach. In the validation sets, the pooled AUROC of DL was 0.908 (95% CI 0.892-0.924; grade: weak), with a sensitivity of 0.87 (95% CI 0.78-0.92) and specificity of 0.93 (95% CI 0.88-0.96). In the validation sets, the model integrating CT-based radiomics with clinical characteristics yielded an AUROC of 0.874 (95% CI 0.852-0.896; grade: weak). The corresponding sensitivity was 0.86 (95% CI 0.77-0.92), and the specificity was 0.75 (95% CI 0.64-0.84). Conversely, for MRI-based models combined with clinical variables, the pooled AUROC was 0.921 (95% CI 0.867-0.979; grade: weak). The corresponding sensitivity and specificity were 0.88 (95% CI 0.81-0.93) and 0.88 (95% CI 0.76-0.94), respectively (Tables 2-3).

HER2 is encoded by the ERBB2 gene. It plays a critical role in the development of various malignancies, including breast, gastric, bladder, ovarian, and lung cancers. HER2-targeted therapies have become first-line treatments for patients with advanced cancers exhibiting HER2 overexpression. Three studies reported on radiomics-based approaches to assess HER2 expression in bladder cancer. Yu et al [55] used MRI radiomics-based ML for noninvasive assessment of HER2. In their test set, an SVM model demonstrated an AUC of 0.712 (95% CI 0.535-0.889), with a sensitivity of 0.857 and a specificity of 0.533. They highlighted the potential value of this approach in cases where patients cannot undergo invasive procedures such as biopsies or diagnostic TUR. Based on contrast-enhanced computed tomography (CE-CT), Peng et al [74] assessed a clinical-radiomics model in evaluating HER2 status in urothelial bladder cancer. This CE-CT-based model exhibited the highest effectiveness in forecasting HER2 status, with an AUC of 0.814 (95% CI 0.642-0.986) in the test set. Wei et al [77] conducted a multicenter study using explainable ML based on CT radiomics to preoperatively diagnose HER2 status in bladder cancer. In their test set, AUCs were 0.803, 0.709, 0.679, 0.794, and 0.815 for LR, SVM, KNN, XGBoost, and RF models, respectively. These studies suggest that radiomics-based ML has promising potential for detecting HER2 expression in bladder cancer.

Ki-67 expression is associated with a poor prognosis and advanced clinicopathological features in cancer. Two single-center studies [35,76] explored radiomics-based prediction of Ki-67 expression in bladder cancer, with the potential to improve prognostic assessment and clinical decision-making. Zheng et al [35] reported on an MRI-based radiomics study. Their SMOTE-LASSO model achieved an AUC of 0.819 (0.658-0.980) in the validation set, with a sensitivity of 0.795 and a specificity of 0.867%. Feng et al [76] developed a radiomics nomogram based on CE-CT, which demonstrated AUCs of 0.836 and 0.887 in their validation set. These findings indicate that radiomics-based ML holds promise for detecting Ki-67 expression in bladder cancer.

Additionally, 2 [37,68] studies reported on the use of radiomics to predict LN staging. Gresser et al [68] used CT radiomics to evaluate LN staging in bladder cancer. Their combined model incorporated manually segmented radiomic features and radiologist assessment. The model achieved an AUC of 0.81 (0.71-0.92), a sensitivity of 0.73 (0.55-0.88), and a specificity of 0.84 (0.76-0.92), respectively, in the test set. However, Starmans et al [37] conducted a multicenter study using preoperative CT radiomics. They found that in patients with cT2-T4aN0-N1M0 muscle-invasive bladder cancer (MIBC), radiomics was not helpful in differentiating pN+ and pN0 disease. This suggests that further research is needed to explore the effectiveness of radiomics-based ML in detecting LN staging in bladder cancer.

Funnel plots were generated to evaluate small-study effects in radiomics-based ML models for detecting muscle invasion and high-grade tumors. Significant small-study effects were observed in both the training and validation sets for assessing muscle invasion (P<.05; Figures S15-S16 in Multimedia Appendix 3), as well as for identifying high-grade tumors (P<.05; Figures S17–S18 in Multimedia Appendix 3).

For the tasks of detecting muscle invasion, meta-regression was performed. Independent variables included case number, model type (LR [Reference] vs other ML), and modeling variables (Radiomics [Reference] vs. Radiomics +clinical). Separate analyses were conducted for models developed on CT or MRI in the training set and validated on CT or MRI in the validation set. The results indicated a significant association between case number and the AUC only in the validation set of MRI-based models for detecting muscle invasion (P<.05). No other variables demonstrated a significant influence (P>.05; Table S2 in Multimedia Appendix 4 and Table S3 in Multimedia Appendix 5). Meta-regression was not conducted for pathological grade due to an insufficient number of studies.

This systematic review incorporated 57 studies [28-84] to evaluate radiomics-based ML for preoperative risk stratification in bladder cancer. The findings indicated a high AUROC for detecting both muscle invasion and high-grade tumors. Additionally, radiomics showed potential for identifying HER2 and Ki-67 expression. However, evaluation via the RQS revealed an overall low methodological quality among the eligible studies. PROBAST-AI assessment revealed that the primary source of bias was in the model evaluation phase, primarily due to small validation set sizes or an absence of external validation. Consequently, the findings of this review should be interpreted with caution. Based on the present findings, radiomics is considered to have application potential. Nevertheless, the current evidence faces significant challenges, including methodological shortcomings and a high ROB, which currently preclude its readiness for clinical translation.

Previous studies have examined the application of radiomics in bladder cancer. Kozikowski et al [10] reviewed the prediction ability of radiomics in muscle-invasive cancer, ultimately including eight articles. The pooled estimated sensitivity and specificity were 82% (95% CI 77%‐86%) and 81% (95% CI 76%‐85%). However, the study did not classify or analyze CT and MRI image sources separately nor attempt to combine different ML approaches with clinical radiomics models. Boca et al [14] reviewed MRI-based radiomics studies in bladder cancer, ultimately including 26 articles with 2991 participants. Radiomics in these studies was primarily used for preoperative prediction of tumor stage or molecular correlations (n=9), preoperative tumor grading (n=13), and prediction of prognosis or response to neoadjuvant therapy (n=4). Most radiomics models incorporated second-order features from filtered images, with quality scores ranging from 8.33% to 52.77% [14]. However, the study only discussed MRI-based radiomics and failed to consider other imaging modalities. The study also did not analyze sensitivity, specificity, or dataset differences, nor did it incorporate a discussion of radiomics combined with clinical features. Building on these prior efforts, the current study conducted a more systematic and comprehensive review of the current state of radiomics-based ML in bladder cancer. The current study also considered different image features, accounted for various image sources, analyzed different modeling approaches, and examined the detection performance across diverse datasets. This provides more comprehensive evidence for future developments.

In the clinical management of bladder cancer, CT often struggles to accurately diagnose flat lesions and prostate-adjacent bladder base lesions, especially in patients with benign prostatic hyperplasia. This difficulty arises from the challenge of distinguishing tumor recurrence from inflammatory wall thickening following intravesical chemotherapy, as well as from scar tissue after TURBT [85]. MRI assessment is a laborious, slice-by-slice process, with its effectiveness depending on the radiologist’s experience [86]. The integration of radiomics and ML with bladder MRI holds promise for improving staging and treatment response assessment. Bladder cancer management guidelines would be enhanced by the integration of MRI into their staging strategies [87]. However, radiomics may experience similar slow progress as molecular biology-based diagnostic and therapeutic techniques. This can be attributed to several factors, such as technical complexities, a lack of validation standards, and inadequate study design (particularly conflating hypothesis generation with hypothesis testing) [11]. Other contributing factors include data overfitting, incomplete reporting of results, and unidentified confounding variables in the datasets used (especially in retrospective datasets). Therefore, as with all biomarker studies, retrospective radiomics studies require validation on independent datasets, ideally from another institution [11].

In clinical practice, MRI has become the most accurate imaging modality for evaluating local invasiveness in bladder cancer. It can be used to assess regional LN involvement and tumor spread to pelvic bones and the upper urinary tract [21,88]. While bladder cancer appears as a soft tissue lesion on CT scans, it is more easily identified as a filling defect during CT urography. Ultrasound is a dynamic imaging modality. It can distinguish bladder cancer from other conditions that appear similar. Its ability to detect blood flow aids in distinguishing the solid tissue of the tumor from blood clots or debris [89]. This meta-analysis reveals that in studies examining muscle invasion, the primary image sources are CT, MRI, and, to a lesser extent, ultrasound. Radiomics combined with clinical features has not consistently demonstrated superior detection performance, potentially due to limited available data. The diagnostic effectiveness of radiomics warrants further investigation. In studies investigating histological grading, the primary image sources are CT and MRI. Only a few studies have explored combining radiomics with clinical features, resulting in limited current evidence.

Traditional ML-based radiomics requires substantial upfront effort, including texture extraction, manual image segmentation, and model construction. This process carries the risk of image data loss and imposes a significant workload. Manual segmentation, in particular, can be influenced by clinical experience, habits, and prior research [86,90]. Consequently, some researchers have explored DL, leveraging its capabilities for staging, grading, automated tumor detection, intelligent segmentation, bladder wall segmentation, and prediction of recurrence, response to chemotherapy, and overall survival. The goal is to improve disease management [58,91-94]. While external validation of intelligent segmentation accuracy is lacking, the potential of DL to reduce workload and automatically interpret images in an intelligent manner is encouraging the development of better, smarter tools. Of the studies included in this review, only a few DL models have demonstrated good performance. Thus, future research should explore DL in addition to traditional methods.

While the models in this study demonstrated relatively high AUROC values, the widespread absence of calibration metrics represents a major limitation. This gap substantially undermines the reliability of these models for practical application [95,96]. Dependence on discriminatory metrics such as AUC, sensitivity, and specificity only reflects the ability of a model to differentiate between positive and negative events. These metrics do not, however, inform on the accuracy of the predicted probabilities—a critical shortfall that poses challenges for safe integration into clinical decision support [97,98]. Consequently, future studies must report comprehensive calibration metrics, including calibration curves, calibration slope, intercept, and the Brier score. Furthermore, decision curve analysis should be performed to evaluate the clinical net benefit across different decision thresholds, a crucial step for assessing clinical utility.

Before initiating this meta-analysis, a prospective registration was completed. However, several adjustments were made during the actual research process. First, to mitigate the risk of missing newly published literature and ensure the completeness of evidence, supplementary database searches were conducted on October 17, 2025. Second, to better reflect the ROB in the included studies, PROBAST-AI was adopted in addition to RQS for the assessment of bias. Third, the GRADE approach was applied to evaluate the certainty of evidence. Fourth, subgroup analyses were carried out based on the type of dataset (training vs validation), imaging source, and type of model. Finally, analyses of small-study effects and meta-regression were also performed. The registered protocol will be updated accordingly.

This meta-analysis is the first comprehensive and systematic review of preoperative risk stratification for bladder cancer, summarizing the available evidence. However, it is subject to the following limitations. First, the present analysis incorporated only 16 studies [29-32,42,47,50,52,58-60,77,78,80,81,83] that used external validation. Among these, 13 investigations focused on muscle invasion, one examined histological grade, one assessed HER2 expression, and one addressed preoperative LN staging. The analytical strategy required subgrouping based on the imaging modality used for model development and the integration of clinical features. Incorporating the data source as a subgroup variable was precluded due to the limited number of multicenter datasets, which would have yielded statistically unreliable comparisons. Furthermore, the validation sets in this study were predominantly created via random sampling. The lack of independent external validation limits the interpretation of the results to some extent. Second, different mathematical models exhibit varying performance in processing images. Although subgroup analyses were conducted to account for these differences, the small number of models within each subgroup may limit the interpretation of our findings. Therefore, the comparative performance of models within these subgroups should be interpreted with caution. Third, evidence for the preoperative identification of LN metastasis and Ki-67 expression is extremely limited. Fourth, while subgroup analyses were performed based on different model types, definitive conclusions cannot be drawn. However, given the paucity of available studies and the small number of articles included in specific model categories, definitive conclusions cannot be drawn. Fifth, the RQS and PROBAST-AI instruments incorporate elements requiring subjective judgment, which may introduce variability into the overall interpretation of the quality assessments. Sixth, a notable limitation of the current evidence base is the infrequent reporting of calibration metrics, such as the calibration-in-the-large, calibration slope, and Brier score. This omission considerably diminishes the clinical applicability of the model performance estimates. Seventh, while all eligible studies confirmed diagnoses pathologically, the transparency regarding the origin of pathological specimens was often insufficient. This lack of detail concerning specimen sourcing is a potential source of bias. Together, these limitations collectively curtail the generalizability of the study findings. Eighth, a notable small-study effect was observed across the included studies. Thus, the findings should be interpreted with caution.

In summary, significant challenges remain prior to the clinical deployment of these models. First, current investigations have not rigorously examined the impact of imaging modality. Variations in acquisition protocols can influence image quality; however, the potential effect of such heterogeneity on modeling outcomes has not been adequately addressed in the existing literature. Second, independent external validation is essential for establishing the generalizability of models. The predominant reliance on internal validation in the eligible studies implies that these findings propose a potential methodological approach rather than one ready for clinical application. Substantial additional evidence is required to confirm clinical utility and generalizability. Third, sample size presents a major constraint. Only a few studies provided a sufficient number of cases, and most encompassed fewer than 200 subjects. Consequently, discussions regarding model robustness necessitate larger, more powerful datasets. Fourth, the image segmentation process depends heavily on manual delineation. This approach is time-consuming and susceptible to variability introduced by the operator’s prior knowledge. Future work should prioritize developing and implementing DL-based automated or semiautomated segmentation techniques.

To facilitate successful clinical translation, future investigations should prioritize the following areas: conducting robust pilot studies, performing reproducibility analyses on different imaging protocols and segmentation methods, ensuring adequate sample sizes, implementing multicenter external validation, comparing diverse ML architectures, and establishing protocols for continuous model updating and refinement.

This is the first systematic review to comprehensively assess radiomics for preoperative risk stratification in bladder cancer. The findings provide evidence to support the development and refinement of future ML-based tools for image analysis. However, due to limitations in the current evidence, such as methodological flaws and a high ROB, low GRADE level, clinical translation is not yet warranted. Future research should standardize radiomics workflows, incorporate multicenter images from diverse geographical regions, and minimize the impact of imaging protocols and pre-processing steps. These measures are essential to advance radiomics toward successful clinical implementation. Such efforts are essential for fully elucidating and validating the potential of radiomics in the noninvasive diagnosis of bladder cancer.