Artificial intelligence advances in cystoscopy and imaging for bladder cancer: a narrative review

Usman Khalid¹, Nikhil Shah¹, Rajesh Kavia², Deepak Batura^2,*
1 Faculty of Medicine, Medical University of Plovdiv, Plovdiv, Bulgaria
2 Department of Urology, London North West University Healthcare NHS Trust, Watford Road, Harrow, London, UK
* Corresponding Author: Deepak Batura. Email: email
(This article belongs to the Special Issue: Advancing the Diagnosis and Treatment of Urological Diseases through Big Data)

Canadian Journal of Urology https://doi.org/10.32604/cju.2026.074820

Received 18 October 2025; Accepted 18 February 2026; Published online 01 April 2026

Download PDF

Abstract

Bladder cancer (BCa) diagnosis relies heavily on cystoscopy and imaging. Both have limited sensitivity and accuracy, particularly for muscle-invasive disease. Artificial intelligence (AI) has emerged as a promising tool for improving detection, grading, and staging by extracting imaging features that exceed human perception. We conducted a narrative review of peer-reviewed, English-language studies published between 2015 and 2025. We identified 75 articles and synthesized data from 35 key studies retrieved via PubMed, Google Scholar, Scopus, and Embase. Data were synthesized narratively, emphasizing diagnostic performance, clinical relevance, and study limitations. In cystoscopy, AI models achieved high accuracy in tumour detection and grading, including carcinoma in situ, and in some studies reduced missed lesions by >20% compared with expert urologists; CystoNet-T reported an average precision of 91.4%. Performance varied across studies, with inconsistencies driven by methodological factors: endoscopic acquisition conditions (illumination quality; white-light vs. enhanced imaging), variation in ground truth and annotation (histopathology vs. expert-labelled frames), dataset size and class imbalance, model architecture and preprocessing pipelines, and validation strategy (internal splits vs. external testing). Most evidence remains limited by retrospective, single-center datasets, which restrict generalizability. In imaging, AI applications showed modality-dependent performance. CT-based radiomic and deep-learning models demonstrated the most consistent improvements for grading, staging, and recurrence prediction across several studies. MRI and radiogenomic models demonstrated proof-of-concept associations between imaging features and molecular profiles. However, clinical readiness is limited by small cohorts and a lack of prospective validation. Results were also affected by scanner heterogeneity and reader dependence. Evidence for PET/CT remains sparse and preliminary. Many approaches outperformed conventional clinical models, including a vision transformer (ViT)-based MRI model for muscle invasiveness prediction (AUC = 0.872). Methodological heterogeneity restricted generalisability. AI shows potential to reduce diagnostic variability and improve initial treatment planning in BCa. Nonetheless, prospective multicentre trials with standardised methodology are essential before clinical integration. In parallel, regulatory approval of AI as medical software, real-time workflow integration within cystoscopy suites and radiology systems, data governance and patient privacy, and ongoing post-deployment performance monitoring must be addressed to ensure safe and effective clinical use.