Bladder cancer is among the 10 most commonly diagnosed cancers worldwide. At initial presentation, approximately 75% of cases are classified as nonmuscle-invasive bladder cancer (NMIBC), while the remainder present as muscle-invasive bladder cancer (MIBC) or metastatic disease.¹ Although NMIBC carries a relatively favorable 5-year survival rate of around 90%, it is marked by high recurrence rates, and up to 20% of cases eventually progress to MIBC, where prognosis declines sharply.

Current guidelines provide risk stratification systems based on clinical and pathologic features to guide treatment decisions.^2,3 However, significant challenges remain: patients within the same risk category often follow divergent disease trajectories, complicating choices regarding adjuvant intravesical therapy, surveillance intensity, and the optimal timing of radical surgery.⁴ Furthermore, there are no biomarkers for prediction of response to various therapies.

In recent years, artificial intelligence (AI) has emerged as a promising tool to address these limitations. By leveraging machine learning and deep learning (DL) approaches, AI enables reproducible, data-driven analyses that can integrate histologic, molecular, and clinical information. These tools have the potential to refine diagnostic accuracy, improve risk stratification, and support personalized treatment strategies.

AI in NMIBC: The Computational Histology AI Platform

A series of studies has introduced the computational histology AI (CHAI) platform, a DL-based tool that analyzes routine pretreatment hematoxylin and eosin slides to predict outcomes in NMIBC (Figure).^5-8 In a large multicenter study including 944 patients with high-risk NMIBC treated with intravesical bacillus Calmette-Guérin (BCG), Lotan et al demonstrated that the CHAI assay stratified patients into distinct prognostic groups.⁵ Those identified as “high risk” had significantly inferior high-grade recurrence-free survival (HR 2.08, P < .0001) and progression-free survival (HR 3.87, P < 0.001) compared with “low-risk” patients. Importantly, the AI-derived signatures provided predictive information beyond standard clinicopathologic variables and outperformed existing European Association of Urology and AUA risk models, especially in challenging subgroups, such as those with high-grade Ta disease, for which existing guidelines often provide limited discrimination.⁸

Beyond initial prognostication, CHAI may also guide management after recurrence. In a pilot study of 52 patients who recurred following an initial course of BCG, application of the CHAI biomarker to recurrent tumors identified those unlikely to benefit from a second induction.⁷ Patients with the biomarker present experienced a markedly higher 6-month recurrence rate (70%) compared with only 22% among those without the biomarker (HR 2.41, 95% CI 1.01-5.75; P < .05). These findings suggest a potential role for CHAI in selecting patients for BCG rechallenge vs early transition to alternative therapies.

Perhaps most notably, recent data indicate that the CHAI biomarker is therapy specific rather than merely prognostic. In a study of 253 patients with high-grade NMIBC treated with either intravesical BCG or sequential gemcitabine/docetaxel, Packiam et al found a significant interaction between treatment type and biomarker status.⁶ Among biomarker-positive patients, 24-month high-grade recurrence-free survival was substantially lower with BCG than with gemcitabine/docetaxel (56% vs 90%; HR 5.4, 95% CI 1.6-18.3; P = .007). By contrast, in biomarker-negative patients, outcomes did not differ between the 2 regimens (HR 1.3, P = .5). The biomarker-treatment interaction was statistically significant (P = .029), supporting its role as a predictive marker of BCG resistance rather than a general marker of adverse biology.

AI in MIBC: Toward Multimodal Biomarkers

While much attention has focused on predictive biomarkers in NMIBC, the management of MIBC presents an equally urgent challenge. Neoadjuvant chemotherapy followed by radical cystectomy remains the standard of care. Yet only about one-third of patients achieve a complete pathologic response, the subgroup most strongly associated with long-term survival. The rest endure chemotherapy toxicity without clear benefit, underscoring the critical need for predictive biomarkers to guide treatment selection.

AI has been investigated in this context. A comprehensive systematic review and meta-analysis by Suartz et al evaluated 12 studies (1523 patients) using AI to predict neoadjuvant chemotherapy response.⁹ In pooled analysis of radiomics-based studies, machine-learning models achieved a specificity of 82% (95% CI 0.72-0.89) and a more modest sensitivity of 62% (95% CI 0.50-0.72). These results highlight the potential of AI to rule out nonresponders but also the limitations of unimodal models.

To address these shortcomings, Bai et al recently developed an interpretable graph-based multimodal late fusion DL framework.¹⁰ Using prospectively collected data from the SWOG S1314 trial, the model integrated 2 complementary data streams: (1) tissue architecture and cell morphology from routine hematoxylin and eosin–stained whole-slide images, and (2) transcriptomic profiles from RNA sequencing. The multimodal model achieved a mean AUC of 0.74 in cross-validation, outperforming each unimodal branch (AUC ∼0.71 for gene expression and 0.67-0.72 for histology alone). Crucially, the authors emphasized interpretability. Through SHAP (Shapley additive explanations)–based analysis, the model identified biologically relevant drivers of response, including alterations in TP63 (tumor protein 63), CCL5 (chemokine ligand 5), and DCN (decorin), as well as histologic correlates such as a higher tumor-stromal ratio in responders. These findings suggest that multimodal AI frameworks can not only improve prediction accuracy but also provide mechanistic insights into tumor biology, moving the field closer to clinically actionable, biologically informed biomarkers in MIBC.

Challenges and Future Directions

The promise of AI in bladder cancer is undeniable, yet several barriers must be overcome before routine clinical adoption. The robustness of any AI model ultimately depends on the quality, scale, and representativeness of its training data. Major hurdles include managing intercohort heterogeneity, reducing algorithmic bias that may arise from imbalanced datasets, and ensuring rigorous external validation across diverse patient populations. The persistent “black box” problem, whereby predictions are made without transparent rationale, continues to undermine clinician confidence.¹¹ In parallel, concerns regarding data privacy, security, and accountability raise important ethical and regulatory questions.¹²

Another challenge is the rapidly evolving therapeutic landscape. Neoadjuvant immunotherapy,¹³ novel intravesical delivery platforms such as TAR-200,¹⁴ and gene-based therapies are shifting standards of care.¹⁵ A biomarker trained solely to predict response to cisplatin-based chemotherapy or BCG risks rapid obsolescence if chemo-immunotherapy or other modalities become the new standard. Thus, future AI models must be designed with adaptability in mind, capable of recalibration as treatment paradigms change.

Finally, generalizability remains a key obstacle, particularly in the domain of digital pathology. Whole-slide image variability, including scanner brand, staining protocols, and segmentation approaches, can compromise reproducibility and limit performance across institutions.¹⁶ To move beyond proof-of-concept, future research should prioritize not only model accuracy but also real-world validation, health-economic evaluation, and prospective clinical trials.

Conflicts of Interest: Dr Lotan is a consultant for Photocure, Astra-Zeneca, Merck, Fergene, Nucleix, Ambu, Seattle Genetics, Virtuoso Surgical, Stimit, Urogen, Vessi Medical, CAPs Medical, Nonagen, Aura Biosciences Inc, Convergent Genomics, Pacific Edge, Pfizer, Phinomics Inc, CG Oncology, Uroviu, Promis Diagnostics, Valar Labs, Uroessentials, NRx Pharmaceuticals, Vesica Health, Janssen, Immunity Bio, Trigone Pharma, and Relmada. No other disclosures were reported.