Artificial Intelligence in Bladder Cancer: Current Applications, Challenges, and Future Perspectives Across Prevention, Diagnosis, Prediction, Treatment, and Supportive Care

1. Introduction

1.1. Global Burden of Bladder Cancer

Bladder cancer remains one of the most common malignancies of the urinary tract and continues to represent a major global health burden [1,2]. It is among the most frequently diagnosed cancers worldwide, accounting for more than half a million new cases annually and causing substantial cancer-related morbidity and mortality [1]. The disease demonstrates a marked male predominance, largely attributable to higher rates of cigarette smoking, occupational exposure to aromatic amines and other industrial carcinogens, and potential hormonal and genetic influences. Bladder cancer incidence also increases significantly with advancing age, with most patients diagnosed after the sixth decade of life [1]. As populations continue to age worldwide and life expectancy increases, the overall burden of bladder cancer is expected to rise further, placing growing pressure on healthcare systems and emphasizing the need for improved strategies for prevention, early detection, and long-term disease management [1,3,4].

Bladder cancer is characterized by substantial clinical and biological heterogeneity [1,3,4]. Approximately 70–75% of newly diagnosed tumors are classified as non–muscle-invasive bladder cancer (NMIBC), encompassing papillary tumors and carcinoma in situ confined to the mucosa or lamina propria [3]. Although patients with NMIBC generally have favorable long-term survival, recurrence is extremely common, and a significant proportion eventually progresses to muscle-invasive bladder cancer (MIBC) [3,17]. MIBC is characterized by invasion into the detrusor muscle and is associated with a markedly increased risk of regional and distant metastasis as well as cancer-specific mortality [4]. Despite major therapeutic advances, including immune checkpoint inhibitors, antibody–drug conjugates, and targeted therapies, advanced and metastatic urothelial carcinoma continue to present significant therapeutic challenges and remain associated with poor long-term outcomes [4,18,19,20]. This broad disease spectrum contributes to highly variable clinical courses and makes individualized patient management particularly challenging.

The burden of bladder cancer extends beyond oncologic outcomes because of its chronic, recurrent nature and the intensive surveillance required throughout a patient's lifetime [2,3]. Patients with NMIBC frequently undergo multiple transurethral resections of bladder tumor (TURBT), repeated courses of intravesical therapy, and lifelong cystoscopic surveillance, often accompanied by urine cytology and periodic upper urinary tract imaging [3,21]. These repeated procedures contribute not only to substantial healthcare expenditures but also to considerable psychological distress, anxiety regarding recurrence, and diminished quality of life. Consequently, bladder cancer remains one of the most expensive malignancies to manage on a per-patient basis among urologic cancers because of continuous surveillance, repeated interventions, and treatment of recurrent or advanced disease [2]. Collectively, these clinical and economic challenges highlight the need for more accurate, efficient, and individualized approaches to bladder cancer care.

1.2. Why Bladder Cancer Is Particularly Suitable for Artificial Intelligence

Bladder cancer is especially well suited for artificial intelligence (AI) applications because its diagnosis and management generate large volumes of multimodal and longitudinal clinical data throughout the patient's disease course [5,9]. Routine clinical evaluation typically includes cystoscopic imaging and video recordings, urine cytology, digital pathology obtained from biopsy or transurethral resection specimens, computed tomography urography, multiparametric magnetic resonance imaging (mpMRI), molecular and genomic profiling, and an expanding array of urinary biomarkers [2,9,10,11]. Unlike many other solid tumors, bladder cancer also requires lifelong surveillance because of its exceptionally high recurrence rate, resulting in rich longitudinal datasets that capture disease evolution, treatment response, and recurrence over time [3].

The integration of imaging, pathology, molecular, and clinical information provides an ideal foundation for computational analysis and precision oncology [5,6]. Machine learning algorithms are particularly effective in this setting because they can analyze high-dimensional datasets, identify complex nonlinear relationships among clinical variables, and uncover imaging or histopathologic features that may be imperceptible to human observers [13,22,23]. Deep learning approaches have demonstrated promising performance in automated lesion detection during cystoscopy, digital pathology interpretation, urine cytology classification, radiomics-based staging, and prediction of recurrence and treatment response [7,8,12,16]. Moreover, repeated surveillance enables continuous updating of individualized risk predictions as new clinical information becomes available, supporting dynamic clinical decision-making over time [5]. Beyond improving diagnostic and prognostic accuracy, AI technologies may also enhance healthcare efficiency by assisting with image interpretation, pathology screening, clinical documentation, scheduling, and patient communication [6]. Together, these unique characteristics make bladder cancer one of the most promising disease models for the implementation of AI-driven precision medicine.

1.3. Overview of Artificial Intelligence in Oncology and Bladder Cancer

Artificial intelligence (AI) encompasses a broad range of computational methods designed to perform tasks that traditionally require human intelligence, including pattern recognition, prediction, decision-making, and data interpretation [6]. Within AI, machine learning (ML) refers to algorithms that automatically learn relationships between input variables and clinical outcomes from existing datasets [6,22]. Supervised learning, the most widely used approach in medical research, trains algorithms using labeled datasets—for example, digital pathology images linked to tumor grade or clinical records associated with recurrence and survival outcomes [13,22]. In contrast, deep learning (DL), a subset of machine learning, employs multilayer artificial neural networks that automatically extract increasingly complex features directly from raw data without extensive manual feature engineering [23,24].

Among deep learning architectures, convolutional neural networks (CNNs) have become the predominant approach for medical image analysis because of their ability to recognize hierarchical spatial features in radiologic images, whole-slide pathology, and endoscopic video sequences [23,24]. These models have demonstrated excellent performance in tumor detection, image segmentation, histopathologic classification, and radiologic staging across multiple cancer types, including bladder cancer [7,8,12]. More recently, transformer-based architectures and foundation models have further expanded AI capabilities by improving image understanding, multimodal learning, and knowledge integration across diverse clinical datasets.

In parallel, natural language processing (NLP) and large language models (LLMs) have emerged as powerful tools for extracting clinically relevant information from electronic health records, generating structured clinical documentation, summarizing scientific literature, supporting multidisciplinary decision-making, and improving patient education [6]. Although clinical implementation remains in its early stages, these technologies have considerable potential to reduce administrative burden and facilitate more efficient clinical workflows.

Another major development is the emergence of multimodal AI, which integrates heterogeneous data sources—including cystoscopic images, radiologic imaging, digital pathology, genomic and molecular profiling, urinary biomarkers, and longitudinal clinical records—within a single predictive framework [5]. Because bladder cancer routinely generates each of these complementary data types throughout diagnosis, treatment, and surveillance, it represents one of the most suitable disease models for multimodal AI applications. Integrating diverse datasets enables more comprehensive diagnostic assessment, individualized risk prediction, treatment selection, and longitudinal disease monitoring than any single data modality alone [5].

1.4. Objectives and Scope of This Review

The rapid evolution of artificial intelligence has stimulated substantial interest in its application to bladder cancer management [6,9]. Although numerous studies have investigated individual applications—including AI-assisted cystoscopy, digital pathology, urine cytology, radiomics, and molecular biomarker analysis—a comprehensive synthesis encompassing the entire continuum of bladder cancer care remains limited [7,8,12,16].

This review provides a comprehensive overview of current AI applications in bladder cancer prevention, diagnosis, prognostic assessment, treatment selection, precision oncology, and survivorship. In addition to summarizing recent technological advances, we critically evaluate the current evidence supporting clinical implementation and discuss the major challenges that continue to limit widespread adoption, including data quality and standardization, external validation, model interpretability, workflow integration, regulatory approval, ethical considerations, and patient privacy [5,6].

Finally, we highlight emerging research directions, including multimodal AI, foundation models, large language models, radiogenomics, dynamic risk prediction, and digital twin technologies, which are expected to play increasingly important roles in precision bladder cancer care over the coming decade [5,6]. By integrating current evidence across multiple disciplines, this review aims to provide clinicians and researchers with a practical and clinically relevant perspective on how AI may reshape bladder cancer diagnosis, treatment, and long-term management.

2. Artificial Intelligence in Prevention and Risk Stratification

2.1. Risk Factor–Based Prediction and Individual Risk Stratification

Artificial intelligence (AI) is increasingly being explored as a tool to improve bladder cancer prevention through more accurate risk stratification and identification of individuals at elevated risk before clinical disease becomes apparent (Lenis et al., 2020; Salzano et al., 2026). Bladder cancer is associated with several well-established risk factors, including cigarette smoking, occupational exposure to aromatic amines and industrial chemicals, arsenic exposure, increasing age, male sex, and family history (Sung et al., 2021; Lenis et al., 2020). Although these risk factors are well recognized, their interactions are highly complex and vary substantially among individuals, making accurate risk estimation difficult using traditional statistical models. Machine learning algorithms offer important advantages by identifying nonlinear relationships among multiple demographic, environmental, lifestyle, and clinical variables and generating individualized risk predictions [6,22].

Recent studies have applied machine learning models to electronic health records, demographic information, occupational exposure histories, and smoking-related variables to improve risk prediction beyond conventional regression-based approaches [5,6]. Because cigarette smoking remains the most important modifiable risk factor for bladder cancer, AI-based models may better quantify individualized risk by incorporating smoking intensity, duration, cessation history, and interactions with occupational or environmental exposures [1,2]. Such models may facilitate targeted smoking cessation interventions, individualized preventive counseling, and earlier diagnostic evaluation of patients presenting with persistent hematuria or additional urothelial cancer risk factors (Lenis et al., 2020). Ultimately, AI-driven risk assessment may enable more personalized prevention strategies than conventional population-based risk models.

Despite these potential benefits, AI applications for bladder cancer prevention remain in an early stage of development [5]. One major limitation is the incomplete and inconsistent documentation of smoking history and occupational or environmental exposures within retrospective clinical datasets, which can substantially reduce predictive accuracy. Furthermore, environmental exposures differ considerably among countries and healthcare systems, limiting model generalizability. Because routine population screening for bladder cancer is not currently recommended, translating AI-based risk prediction into practical clinical screening strategies remains challenging [1]. Future multicenter prospective studies incorporating standardized exposure assessment and longitudinal follow-up will be essential for improving model robustness and clinical applicability.

2.2. Population-Level Epidemiologic Modeling and Public Health Applications

Beyond individual risk prediction, AI has considerable potential to support bladder cancer prevention at the population level through advanced epidemiologic modeling and public health surveillance [5,6]. Classical epidemiologic investigations have demonstrated strong associations between bladder cancer and occupational exposure in industries such as dye manufacturing, rubber production, leather processing, aluminum production, and petrochemical manufacturing [1,2]. However, these exposure patterns frequently involve multiple interacting variables accumulated over decades, making conventional statistical analysis challenging. Machine learning methods can efficiently analyze large heterogeneous datasets and identify complex exposure patterns and high-risk occupational clusters that may remain undetected using traditional epidemiologic approaches [6,22].

AI-driven epidemiologic models may identify geographic regions and occupational populations with elevated disease burden, thereby facilitating targeted prevention strategies and more efficient allocation of healthcare resources [6]. These approaches may also assist public health agencies in forecasting disease incidence, optimizing smoking cessation programs, strengthening occupational safety regulations, and prioritizing high-risk populations for early diagnostic evaluation [5]. Moreover, integration of large healthcare databases may improve understanding of disparities in bladder cancer incidence and outcomes related to socioeconomic status, environmental exposure, healthcare accessibility, and geographic variation [2].

Nevertheless, several important challenges remain. Large administrative and population databases often contain incomplete exposure histories, inconsistent coding practices, heterogeneous diagnostic criteria, and variable follow-up duration, all of which may introduce bias into predictive models. Geographic differences in healthcare access and diagnostic intensity may further influence observed incidence patterns. In addition, ethical concerns related to patient privacy, data security, algorithmic fairness, and responsible use of large-scale healthcare databases must be carefully addressed before widespread implementation [6]. Despite these limitations, population-level AI modeling represents a promising strategy for advancing bladder cancer prevention and public health planning.

2.3. AI-Assisted Biomarker Analysis and Precision Prevention

The rapid development of urinary biomarkers has created new opportunities for AI-assisted bladder cancer prevention and early detection [10,11]. Because urothelial tumors continuously shed malignant cells, nucleic acids, proteins, metabolites, and extracellular vesicles into urine, urinary biomarker testing provides a convenient, noninvasive platform for individualized risk assessment and disease surveillance [2,10]. Biomarkers currently under investigation include cell-free DNA, circulating tumor DNA (ctDNA), DNA methylation signatures, urinary exosomal RNA, proteomic profiles, metabolomic biomarkers, and combinations of molecular assays with conventional urine cytology [9,10]. Because these datasets are high dimensional and biologically complex, they are particularly well suited for machine learning analysis [22].

Rather than relying on individual biomarkers or predefined diagnostic thresholds, AI algorithms can integrate multiple molecular signals to generate comprehensive individualized risk scores [5,6]. Such multimodal biomarker models may improve identification of early urothelial carcinogenesis and facilitate personalized surveillance strategies, particularly among individuals with heavy smoking histories, occupational exposure to carcinogens, or previous urothelial carcinoma [2]. As urinary biomarker technologies continue to mature, AI-assisted models may improve selection of patients requiring early cystoscopy while reducing unnecessary invasive procedures among lower-risk individuals [9,10].

Despite encouraging preliminary results, biomarker-guided prevention strategies remain largely investigational. Many promising biomarkers have demonstrated favorable performance in early validation studies but still lack large prospective multicenter trials confirming reproducibility, clinical utility, and cost-effectiveness [9,10]. False-positive findings may also increase patient anxiety and lead to unnecessary cystoscopy or additional diagnostic testing. Furthermore, laboratory standardization, assay reproducibility, regulatory approval, and accessibility remain important barriers to widespread clinical implementation. Future integration of urinary biomarkers with clinical risk factors, imaging, and longitudinal surveillance data is expected to further improve predictive performance and support precision prevention strategies.

2.4. Current Limitations and Future Directions in Prevention

Although AI has considerable potential to improve bladder cancer prevention, its current clinical application remains relatively limited [5]. Compared with diagnostic imaging and digital pathology, prevention-focused AI models are supported by relatively few prospective studies and limited external validation. Most published models rely on retrospective datasets with relatively small sample sizes, heterogeneous patient populations, and inconsistent exposure documentation, increasing the risk of overfitting and limiting generalizability [5,6]. Variability in biomarker platforms, laboratory methods, and environmental exposure assessment further complicates comparison among studies.

Despite these challenges, prevention remains one of the most promising future directions for AI in bladder cancer. Continued development of large multicenter databases, standardized biomarker assays, harmonized exposure assessment, and integration of clinical, environmental, molecular, and imaging data are expected to improve predictive accuracy and model robustness [5]. Ultimately, AI may facilitate precision prevention by identifying individuals who would benefit most from intensified surveillance or early intervention rather than relying exclusively on broad population-based screening strategies [2]. Such personalized approaches have the potential to improve early detection, reduce unnecessary diagnostic procedures, optimize resource allocation, and contribute to more efficient and cost-effective bladder cancer prevention.

3. Artificial Intelligence in Diagnosis

3.1. AI-Assisted Cystoscopy

Diagnosis represents one of the most mature and rapidly evolving applications of artificial intelligence (AI) in bladder cancer. Cystoscopy remains the gold standard for tumor detection, diagnosis, and surveillance and serves as the cornerstone of both initial evaluation and long-term follow-up [2,3]. However, cystoscopic interpretation is inherently operator dependent and may be influenced by tumor size, morphology, inflammation, bleeding, poor visualization, and endoscopist experience. Small papillary tumors, flat carcinoma in situ (CIS), and subtle mucosal abnormalities are particularly difficult to detect, while incomplete identification of tumor margins during transurethral resection of bladder tumor (TURBT) may contribute to residual disease and subsequent recurrence [3].

Recent advances in deep learning, particularly convolutional neural networks (CNNs), have demonstrated encouraging performance in real-time cystoscopic image analysis [25,26]. AI systems can automatically detect papillary tumors, delineate suspicious mucosal lesions, and assist with lesion localization during both white-light and enhanced cystoscopy [8,25]. Several algorithms generate probability heat maps or visual overlays that highlight regions suspicious for malignancy, potentially improving lesion detection and facilitating more complete tumor resection [7,8]. Because bladder cancer requires lifelong surveillance, AI-assisted cystoscopy may also standardize image documentation, improve longitudinal comparison between surveillance examinations, and reduce interobserver variability [9].

Despite these advances, several important challenges remain. Most published AI models have been developed using retrospective single-center datasets, limiting generalizability across different endoscopic systems, imaging protocols, and patient populations [9]. False-positive findings related to inflammation, previous intravesical therapy, or postoperative scarring may increase unnecessary biopsies and reduce clinician confidence. Furthermore, successful implementation in routine clinical practice will require seamless integration into existing endoscopic platforms while maintaining real-time performance during TURBT. Nevertheless, AI-assisted cystoscopy remains one of the most clinically advanced AI applications in bladder cancer and is expected to play an increasingly important role in improving diagnostic accuracy and surgical quality.

3.2. Artificial Intelligence in Urine Cytology and Urinary Diagnostics

Urine cytology remains an important noninvasive diagnostic and surveillance tool for bladder cancer because of its excellent specificity for high-grade urothelial carcinoma and carcinoma in situ [2]. However, sensitivity for low-grade tumors remains limited, and interpretation is subject to substantial interobserver variability. Diagnostic accuracy may also be affected by specimen quality, inflammatory changes, urinary tract instrumentation, and degenerative cellular alterations [10,11].

Recent developments in digital cytology combined with deep learning have substantially improved automated analysis of urine specimens [10,11]. CNN-based algorithms are capable of identifying atypical urothelial cells, classifying malignant versus benign specimens, and prioritizing suspicious slides for expert review. Rather than replacing cytopathologists, these systems function as decision-support tools that improve diagnostic consistency, reduce workload, and increase laboratory efficiency.

AI has also been increasingly integrated with urinary molecular biomarkers, including DNA mutation profiles, DNA methylation signatures, exosomal RNA, circulating tumor DNA (ctDNA), extracellular vesicles, and proteomic biomarkers [10,11]. Combining cytology with molecular biomarkers enables multimodal diagnostic models that may outperform individual biomarkers alone and improve early detection, particularly among patients with equivocal cytology or persistent microscopic hematuria.

Despite encouraging preliminary findings, several barriers remain before widespread clinical implementation. Variability in specimen processing, slide digitization, staining protocols, and image quality may significantly affect algorithm performance. Moreover, most published studies remain retrospective and have been developed using relatively small datasets without robust external validation [9]. Future multicenter prospective studies with standardized workflows will be essential for establishing the clinical utility of AI-assisted urinary diagnostics.

3.3. Artificial Intelligence in Radiology and Radiomics

Medical imaging plays a central role in the diagnosis, staging, and treatment planning of bladder cancer [2]. Computed tomography urography (CTU) remains the standard imaging modality for evaluating hematuria and upper urinary tract pathology, whereas multiparametric magnetic resonance imaging (mpMRI) has become increasingly important for local staging and assessment of muscle invasion through the Vesical Imaging-Reporting and Data System (VI-RADS) [2][29]. Nevertheless, radiologic interpretation remains subjective and is influenced by reader experience and subtle imaging findings.

Radiomics has emerged as one of the most promising AI applications in bladder cancer imaging [14,15,30]. By extracting hundreds of quantitative imaging features—including texture, shape, intensity, and enhancement characteristics—from routine CT and MRI examinations, radiomics can identify imaging biomarkers that are not readily appreciated by the human eye [14,15,31]. Machine learning algorithms subsequently analyze these features to differentiate tumor stage, predict muscle invasion, estimate tumor aggressiveness, and assess recurrence risk [32,33,34,35,36,37]. Several recent studies have demonstrated that radiomics combined with deep learning significantly improves prediction of muscle-invasive bladder cancer compared with conventional imaging assessment alone [38,39,40,41,42,43]. Furthermore, integrating radiomic features with clinical variables and pathological findings may further enhance individualized risk stratification and treatment planning. Prospective studies have also demonstrated the clinical value of VI-RADS for predicting muscle invasion, providing an important foundation for AI-assisted MRI interpretation [31]. More recently, MRI radiomics has also shown promise for monitoring response to neoadjuvant chemotherapy and predicting treatment outcomes [44].

Despite promising results, radiologic AI continues to face several technical challenges. Differences in scanner manufacturers, acquisition protocols, reconstruction algorithms, and feature extraction methods limit reproducibility across institutions [16]. Most published models remain retrospective and require large multicenter prospective validation before routine clinical implementation. Standardization of imaging acquisition and radiomic analysis will therefore be essential for successful clinical translation.

3.4. Artificial Intelligence in Digital Pathology

Histopathologic examination remains the definitive diagnostic standard for bladder cancer because it determines tumor grade, invasion depth, histologic subtype, and other key prognostic features that directly influence treatment decisions [2,3]. However, interpretation may be challenging in borderline lesions, carcinoma in situ, and uncommon histologic variants, resulting in clinically significant interobserver variability.

Digital pathology has become one of the fastest-growing applications of AI in urothelial carcinoma [12]. Whole-slide imaging combined with deep learning enables automated tumor detection, grading, assessment of invasion, identification of lymphovascular invasion, and recognition of variant histologic patterns [12,13,45]. Weakly supervised learning approaches have further reduced the need for extensive manual annotation while maintaining high diagnostic performance. Beyond conventional morphologic assessment, AI has demonstrated the ability to predict molecular subtypes, genomic alterations, and clinically relevant biological pathways directly from digitized histopathology slides [12,46,47,48]. These emerging applications may eventually complement or reduce the need for additional molecular testing while supporting precision oncology.

Current limitations include variability in tissue fixation, staining protocols, scanner characteristics, and slide preparation across pathology laboratories. The high cost of digital slide annotation and the need for continued pathologist oversight also remain important barriers to implementation. Nevertheless, digital pathology represents one of the most promising AI applications in bladder cancer and is expected to become an integral component of future diagnostic workflows.

3.5. Multimodal Diagnostic Integration

A major emerging direction in bladder cancer diagnosis is the development of multimodal AI, in which cystoscopic images, radiologic imaging, digital pathology, urinary biomarkers, genomic profiling, and longitudinal clinical information are integrated into a unified predictive framework [5]. Rather than evaluating each data source independently, multimodal AI leverages complementary information to generate more accurate and personalized diagnostic assessments.

Such integrated systems may be particularly valuable in challenging clinical scenarios, including carcinoma in situ, equivocal imaging findings, recurrent disease following intravesical Bacillus Calmette–Guérin (BCG) therapy, and discordant pathology or biomarker results [10,12,14]. By combining heterogeneous clinical data within a single predictive model, multimodal AI has the potential to improve diagnostic accuracy, reduce uncertainty, support multidisciplinary decision-making, and facilitate individualized treatment planning.

Although significant challenges remain, including data harmonization, interoperability, regulatory approval, and prospective multicenter validation, multimodal AI represents one of the most promising developments in precision bladder cancer diagnostics. Continued collaboration among clinicians, engineers, pathologists, and data scientists will be essential for translating these technologies into routine clinical practice [5,6].

Table 1 below is a summary of current applications of artificial intelligence (AI) in the diagnosis of bladder cancer. The table highlights the major diagnostic modalities, including cystoscopy, urine cytology, radiology/radiomics, digital pathology, and multimodal AI, together with their principal data sources, commonly used AI approaches, clinical applications, and current limitations. Although many AI models have demonstrated promising diagnostic performance, prospective multicenter validation, standardized data acquisition, and seamless integration into clinical workflows remain necessary before widespread clinical implementation.

4. Artificial Intelligence in Prognosis and Prediction

4.1. Recurrence Prediction in Non–Muscle-Invasive Bladder Cancer

Prediction of disease recurrence remains one of the greatest clinical challenges in non–muscle-invasive bladder cancer (NMIBC). Although long-term survival is generally favorable, recurrence occurs in up to 50–70% of patients, necessitating repeated transurethral resection of bladder tumor (TURBT), intravesical therapy, and lifelong cystoscopic surveillance [3,17]. Widely used clinical prediction tools, including the European Organisation for Research and Treatment of Cancer (EORTC) and European Association of Urology (EAU) risk models, provide useful estimates of recurrence risk but rely on a limited number of clinicopathologic variables and may not adequately capture complex interactions among patient characteristics, tumor biology, and treatment history [3,17].

Machine learning offers important advantages by integrating a broader spectrum of clinical variables, including tumor size, grade, multiplicity, prior recurrence history, smoking status, intravesical therapy, pathological findings, and imaging characteristics [5]. Unlike conventional statistical models, machine learning algorithms can identify nonlinear relationships and complex interactions among predictors, enabling more individualized estimates of recurrence risk [6,22]. Recent MRI-based radiomics studies have further demonstrated that quantitative imaging features extracted from multiparametric MRI can significantly improve prediction of recurrence compared with traditional clinicopathologic models [33,34,45]. More accurate recurrence predictions could facilitate risk-adapted surveillance schedules by reducing unnecessary cystoscopic examinations in low-risk patients while enabling earlier detection of recurrent disease in individuals at higher risk.

Despite encouraging results, several important limitations remain. Most published prediction models have been developed retrospectively using single-center datasets and have undergone limited external validation [5]. Variability in surgical quality, pathological interpretation, intravesical treatment protocols, and follow-up schedules further limits model generalizability across institutions. Nevertheless, recurrence prediction remains one of the most clinically relevant applications of AI in bladder cancer and is likely to become increasingly important as multicenter datasets and prospective validation studies continue to expand.

4.2. Progression Prediction and Identification of Aggressive Disease

Progression from NMIBC to muscle-invasive bladder cancer (MIBC) or metastatic disease represents a critical turning point in bladder cancer management because it is associated with substantially worse survival outcomes and often necessitates radical treatment [3,4]. Although conventional clinicopathologic models provide useful prognostic information, accurately identifying patients at highest risk for progression remains challenging.

Artificial intelligence has the potential to improve progression prediction by integrating radiologic, histopathologic, molecular, and clinical information within a unified predictive framework (Salzano et al., 2026). Radiomics studies have demonstrated that quantitative MRI features can predict muscle invasion and aggressive tumor behavior before surgery [35,36,41]. Similarly, AI-assisted digital pathology can identify subtle morphologic features associated with lymphovascular invasion, variant histology, and poor clinical outcomes that may not be readily appreciated during routine microscopic examination [12,13]. Integration of genomic and transcriptomic information, including consensus molecular subtypes and immune-related biomarkers, may further improve individualized risk stratification [47,48,49,50].

More accurate prediction of disease progression could facilitate earlier treatment intensification, including consideration of radical cystectomy, neoadjuvant systemic therapy, or enrollment in clinical trials before irreversible disease progression occurs [4]. However, challenges related to model reproducibility, data harmonization, and clinical acceptance remain significant, and prospective multicenter validation will be essential before routine clinical implementation.

4.3. Survival Prediction and Outcome Modeling

Artificial intelligence has also been increasingly applied to prediction of long-term oncologic outcomes, including recurrence-free survival, progression-free survival, cancer-specific survival, and overall survival [5]. Because these outcomes are influenced by numerous interacting clinical, pathological, radiologic, molecular, and treatment-related variables, conventional regression models may inadequately capture their complexity.

Machine learning algorithms can integrate demographic characteristics, pathological findings, imaging features, molecular biomarkers, treatment history, and comorbidities to generate individualized survival predictions [6,22,51]. Such models may improve patient counseling, facilitate individualized treatment planning, and assist clinicians in balancing therapeutic benefit against treatment-related morbidity, particularly among patients undergoing radical cystectomy or systemic therapy [4].

Nevertheless, survival prediction remains challenging because outcomes are substantially influenced by treatment heterogeneity, evolving therapeutic strategies, patient frailty, and competing causes of mortality. Furthermore, many currently available AI models have undergone only internal validation and lack robust prospective evaluation across diverse healthcare systems [5]. Continued refinement using large multicenter cohorts and standardized reporting frameworks will be required before widespread clinical adoption.

4.4. Prediction of Treatment Response

Therapeutic response varies considerably among patients with bladder cancer, particularly following intravesical Bacillus Calmette–Guérin (BCG) therapy and systemic treatment for advanced urothelial carcinoma [3,4]. Early identification of patients unlikely to respond to a given treatment remains an important objective because it may facilitate timely treatment modification and reduce unnecessary toxicity.

AI-based predictive models integrate clinical variables, radiologic imaging, digital pathology, genomic profiling, and immune-related biomarkers to estimate the probability of treatment response [45,52]. In NMIBC, these models may identify patients who are unlikely to benefit from intravesical BCG therapy and who may therefore require earlier radical cystectomy or alternative bladder-preserving strategies [3,17]. In advanced disease, integration of molecular subtype classification, fibroblast growth factor receptor (FGFR) alterations, programmed death ligand 1 (PD-L1) expression, and immune microenvironment characteristics may improve prediction of response to immune checkpoint inhibitors, targeted therapies, and antibody–drug conjugates [20,47,49,53,54,55].

Despite promising preliminary results, most treatment-response prediction models remain investigational. Tumor evolution during therapy, spatial and temporal heterogeneity, and changing immune microenvironments introduce additional complexity that remains difficult to model accurately. Large prospective clinical studies incorporating serial molecular and imaging assessments will be necessary to validate these approaches before routine clinical implementation.

4.5. Dynamic Risk Assessment and Longitudinal Surveillance

One of the most promising future applications of AI in bladder cancer is dynamic risk assessment during longitudinal follow-up [5]. Unlike static prediction models generated at the time of diagnosis, dynamic AI systems continuously update individualized risk estimates as new clinical information—including cystoscopic findings, pathology reports, urinary biomarkers, imaging studies, and treatment history—becomes available [6].

Because bladder cancer requires lifelong surveillance with repeated clinical assessments, dynamic prediction models are particularly well suited to this disease [3]. AI algorithms may continuously refine estimates of recurrence and progression risk, thereby enabling adaptive surveillance schedules that reduce unnecessary cystoscopy in low-risk patients while intensifying monitoring among individuals at higher risk. Such approaches have the potential to improve early detection of recurrence, reduce healthcare costs, optimize allocation of clinical resources, and minimize patient burden associated with frequent invasive procedures.

Although dynamic risk prediction remains at an early stage of development, it aligns closely with the broader goals of precision oncology and learning healthcare systems. Future integration of multimodal clinical data, real-time electronic health records, wearable devices, urinary biomarkers, and genomic information may ultimately enable personalized surveillance strategies that evolve continuously throughout each patient's disease course [5,6].

5. Artificial Intelligence in Treatment and Precision Oncology

5.1. Artificial Intelligence in Intravesical Therapy

Artificial intelligence (AI) is increasingly being investigated to support treatment selection for patients with non–muscle-invasive bladder cancer (NMIBC), particularly those receiving intravesical therapy [3,5]. Intravesical Bacillus Calmette–Guérin (BCG) remains the standard treatment for intermediate- and high-risk NMIBC because of its proven ability to reduce recurrence and disease progression [56,57]. Nevertheless, treatment response varies considerably, and a substantial proportion of patients experience recurrence or progression despite adequate BCG therapy [3,17].

Machine learning models have been developed to predict response to intravesical therapy by integrating clinicopathologic variables, including tumor size, grade, multiplicity, prior recurrence history, presence of carcinoma in situ, smoking history, and molecular biomarkers [5,21]. Such models may identify patients who are most likely to benefit from BCG while recognizing those at increased risk of early treatment failure. Earlier identification of BCG-unresponsive disease could facilitate timely treatment intensification, alternative intravesical therapies, systemic immunotherapy, or consideration of early radical cystectomy [3,57].

Despite encouraging preliminary results, most AI-based treatment prediction models remain retrospective and have been developed using heterogeneous treatment protocols. Variability in host immune responses, BCG strains, maintenance schedules, and surveillance strategies further complicates model generalizability. Prospective multicenter validation and standardized reporting will be essential before AI-guided treatment selection can be incorporated into routine clinical practice.

5.2. Surgical Planning and Perioperative Decision-Making

Artificial intelligence also has considerable potential to improve both endoscopic and surgical management of bladder cancer [5]. Transurethral resection of bladder tumor (TURBT) remains the cornerstone of diagnosis and treatment for NMIBC, whereas radical cystectomy remains the standard treatment for muscle-invasive disease and selected patients with high-risk NMIBC [3,4].

AI-assisted cystoscopy may improve lesion detection, define tumor boundaries more accurately, and facilitate more complete tumor resection during TURBT [7,8]. More complete resection may reduce residual tumor burden and potentially decrease recurrence rates. In radical cystectomy, machine learning algorithms have been applied to predict lymph node involvement, perioperative complications, postoperative morbidity, hospital readmission, and long-term oncologic outcomes, thereby supporting individualized preoperative risk assessment and patient counseling [5].

AI has also been explored for optimizing urinary diversion planning by integrating patient age, renal function, comorbidities, functional status, and oncologic characteristics to assist in selecting between continent and incontinent urinary diversion. However, surgical outcomes remain strongly influenced by surgeon experience, institutional case volume, and perioperative care pathways, factors that are difficult to fully capture using predictive models alone. Consequently, prospective validation and integration into clinical workflows remain necessary before routine adoption.

5.3. Systemic Therapy and Precision Oncology

The therapeutic landscape of advanced urothelial carcinoma has evolved rapidly with the introduction of immune checkpoint inhibitors, targeted therapies, and antibody–drug conjugates [4,18,20,54,58]. While these advances have substantially improved patient outcomes, they have also increased the complexity of treatment selection.

Artificial intelligence offers a promising approach to precision oncology by integrating genomic, transcriptomic, radiologic, histopathologic, and clinical information to identify biomarkers associated with therapeutic response or resistance [5,50]. Molecular alterations such as fibroblast growth factor receptor (FGFR) mutations, consensus molecular subtypes, programmed death ligand 1 (PD-L1) expression, and characteristics of the tumor immune microenvironment may all contribute to individualized treatment selection [46,47,48,49,53]. Radiomics and digital pathology further provide quantitative biomarkers capable of predicting immune checkpoint response, PD-L1 expression, molecular subtype, and therapeutic resistance [48,59].

Despite substantial promises, these precision oncology approaches remain largely investigational. Intratumoral heterogeneity, clonal evolution during treatment, and differences among biomarker platforms continue to limit reproducibility and widespread implementation. Future prospective clinical trials incorporating standardized molecular testing and multimodal AI models will be critical for translating these technologies into routine clinical practice.

5.4. Clinical Trial Matching and Clinical Decision Support

Artificial intelligence may also substantially improve clinical trial enrollment by automating identification of eligible patients based on tumor stage, molecular alterations, prior therapies, organ function, and other eligibility criteria [6]. Traditional manual screening of trial eligibility is labor intensive and frequently delays enrollment into potentially beneficial clinical studies.

Natural language processing (NLP) and large language models (LLMs) can automatically extract structured information from electronic health records, pathology reports, radiology reports, and clinical notes to facilitate efficient trial matching and support evidence-based clinical decision-making [6]. These technologies may also assist clinicians by summarizing rapidly expanding oncology literature, generating concise patient summaries, and supporting multidisciplinary tumor board discussions.

Nevertheless, AI-assisted clinical decision support should complement rather than replace clinician judgment. Model transparency, explainability, regulatory oversight, and rigorous quality assurance remain essential to ensure patient safety, minimize algorithmic bias, and maintain clinician confidence before widespread implementation.

6. Artificial Intelligence in Supportive Care and Survivorship

6.1. Patient Education and Shared Decision-Making

Bladder cancer management often involves prolonged surveillance, repeated procedures, and complex treatment decisions that may be difficult for patients to fully understand [2,3]. Artificial intelligence has the potential to improve patient education by translating complex medical information into personalized, accessible language and helping patients better understand their diagnosis, prognosis, treatment options, and expected outcomes [6].

AI-powered educational tools may also facilitate shared decision-making, particularly when discussing bladder-preserving strategies versus radical cystectomy, urinary diversion options, or systemic treatment choices. Improved communication may reduce uncertainty, enhance patient engagement, and promote informed decision-making throughout the disease course.

6.2. Symptom Monitoring and Longitudinal Follow-Up

Patients with bladder cancer frequently experience urinary symptoms, hematuria, postoperative complications, and treatment-related adverse events that require ongoing monitoring [3]. AI-assisted remote monitoring platforms and digital health technologies may facilitate earlier recognition of clinically significant symptoms, allowing prompt clinical evaluation and intervention [6].

Automated reminder systems and intelligent follow-up platforms may also improve adherence to surveillance schedules, particularly among patients with NMIBC who require lifelong cystoscopic surveillance. Continuous symptom monitoring between scheduled clinic visits has the potential to improve patient safety while reducing unnecessary hospital visits and delays in identifying disease recurrence.

6.3. Clinical Workflow and Care Coordination

Beyond direct patient care, AI has considerable potential to improve healthcare efficiency by automating clinical documentation, summarizing pathology reports, generating clinical notes, and assisting with administrative workflows [6]. These technologies may reduce clinician workload and facilitate communication among urologists, oncologists, radiologists, pathologists, and primary care providers, thereby improving multidisciplinary coordination throughout the patient journey.

Despite these potential advantages, successful implementation will require careful attention to patient privacy, cybersecurity, algorithmic bias, digital literacy, and continued clinician oversight. AI-generated recommendations should be viewed as clinical decision-support tools rather than autonomous decision-makers. With continued technological development and appropriate regulatory oversight, AI is expected to become an increasingly valuable component of long-term survivorship care, improving efficiency while maintaining high-quality, patient-centered management.

7. Current Challenges and Limitations

Despite the remarkable progress achieved in recent years, several important challenges continue to limit the widespread clinical implementation of artificial intelligence (AI) in bladder cancer care, most published AI studies remain retrospective and are based on relatively small, single-center datasets, increasing the risk of overfitting and limiting reproducibility across diverse patient populations [26,60]. Furthermore, substantial heterogeneity in imaging acquisition protocols, cystoscopic equipment, pathological specimen preparation, staining techniques, and urinary biomarker platforms complicates model development and reduce generalizability across institutions.

Another major challenge is the limited external validation of currently available AI models. Many algorithms demonstrate excellent performance during internal validation but show reduced accuracy when applied to independent datasets from different healthcare systems or geographic regions [5]. Standardized multicenter prospective validation studies and harmonized reporting guidelines will therefore be essential before AI systems can be routinely incorporated into clinical practice. Model interpretability also remains a critical barrier to clinical adoption. Many deep learning algorithms function as "black-box" systems, making it difficult for clinicians to understand how predictions are generated or to assess their reliability in individual patients [6]. Improving explainability through visualization techniques, transparent model design, and clinician-friendly decision-support interfaces will be important for increasing physician confidence and facilitating clinical acceptance.

Additional concerns include algorithmic bias, patient privacy, cybersecurity, regulatory approval, medico-legal responsibility, and integration into electronic health record systems [6]. AI models trained using nonrepresentative populations may inadvertently perpetuate healthcare disparities if applied broadly without careful evaluation. Furthermore, successful implementation will require close collaboration among clinicians, engineers, regulatory agencies, and healthcare organizations to ensure patient safety, ethical governance, and equitable access to AI technologies.

Addressing these technical, clinical, ethical, and regulatory challenges will be essential before AI can be reliably integrated into routine bladder cancer management. Future studies should prioritize prospective multicenter validation, standardized reporting, transparent model development, and demonstration of meaningful clinical benefit rather than focusing solely on algorithmic performance.

8. Future Directions

Artificial intelligence is expected to play an increasingly important role in bladder cancer management over the coming decade as computational methods become more accurate, interpretable, and integrated into routine clinical workflows [5,6]. One of the most important future directions is the development of multimodal AI, which combines cystoscopic imaging, radiologic examinations, digital pathology, urinary biomarkers, genomic profiling, and longitudinal clinical data within unified predictive frameworks [45,48,59]. Such integrated models have the potential to provide more comprehensive diagnostic assessment, individualized risk prediction, treatment selection, and surveillance planning than any single data modality alone.

Another promising area is dynamic risk modeling, in which recurrence, progression, and treatment-response predictions are continuously updated as new clinical information becomes available [6]. Because bladder cancer requires lifelong surveillance with repeated cystoscopy, imaging, pathology, and biomarker assessment, it represents an ideal disease for adaptive prediction models that evolve throughout the patient's clinical course. Such systems may facilitate individualized surveillance schedules, improve early detection of recurrence, reduce unnecessary invasive procedures, and optimize healthcare resource utilization.

Rapid advances in large language models (LLMs) and natural language processing are also expected to transform clinical practice. These technologies may assist with automated documentation, extraction of clinically relevant information from electronic health records, patient education, literature summarization, multidisciplinary tumor board preparation, and clinical trial matching. As these systems mature, they are likely to become increasingly valuable clinical decision-support tools rather than replacements for physician expertise.

Looking further ahead, the concept of patient-specific digital twins represents one of the most exciting future directions in precision oncology. Digital twins integrate clinical characteristics, imaging, pathology, genomic information, biomarker data, and longitudinal follow-up into individualized computational models capable of simulating disease progression and predicting therapeutic response before treatment is initiated. Although still largely theoretical, this approach has the potential to transform personalized bladder cancer management by enabling simulation-based treatment planning and adaptive precision medicine [5,6].

Ultimately, successful translation of AI into routine clinical care will depend not only on technological innovation but also on rigorous prospective validation, international collaboration, standardized data sharing, transparent regulatory frameworks, and careful integration into existing healthcare systems.

Table 3 below is an overview of the current maturity, major limitations, and future research priorities for artificial intelligence applications across the bladder cancer care continuum. While AI has achieved the greatest clinical maturity in diagnostic imaging and digital pathology, substantial challenges remain in external validation, data standardization, model interpretability, regulatory approval, and clinical integration. Continued development of multimodal AI, large prospective datasets, foundation models, and precision oncology applications is expected to accelerate translation into routine clinical practice.

9. Conclusions

Artificial intelligence has emerged as one of the most rapidly advancing fields in bladder cancer research and clinical oncology. Current applications are most mature in diagnostic imaging, AI-assisted cystoscopy, digital pathology, urine cytology, radiomics, and prediction of recurrence and disease progression, where AI has demonstrated considerable potential to improve diagnostic accuracy, reduce interobserver variability, and facilitate individualized risk stratification [7,8,10,12].

Beyond diagnosis, AI applications are expanding rapidly into treatment selection, precision oncology, surgical planning, prognostic modeling, survivorship care, and clinical decision support. Integration of radiologic, pathologic, molecular, genomic, and longitudinal clinical information offers unprecedented opportunities to personalize treatment strategies and optimize patient outcomes [47,50].

Nevertheless, important challenges remain before AI can be fully incorporated into routine bladder cancer care. Limitations related to data quality, external validation, model interpretability, algorithmic fairness, regulatory approval, and clinical workflow integration continue to represent significant barriers to implementation [5,6]. Addressing these challenges through rigorous prospective clinical studies, standardized reporting, transparent model development, and multidisciplinary collaboration will be essential for ensuring safe and effective clinical translation.

Bladder cancer represents an exceptionally suitable disease model for artificial intelligence because of its multimodal diagnostic pathways, molecular heterogeneity, and lifelong surveillance requirements. As AI technologies continue to evolve and mature, they are expected to become integral components of precision bladder cancer care, supporting earlier diagnosis, more accurate risk prediction, individualized treatment selection, improved survivorship, and ultimately better long-term patient outcomes.