This study presents the first systematic review and meta-analysis of artificial intelligence (AI)-assisted cystoscopic diagnosis in bladder cancer (BC) and interstitial cystitis/bladder pain syndrome (IC/BPS). The primary focus is on evaluating diagnostic performance for carcinoma in situ (CIS) versus Hunner lesions, two flat, erythematous bladder lesions that are visually difficult to distinguish using conventional cystoscopy.
Although recent studies have shown promising results for AI in detecting bladder tumors, its ability to differentiate neoplastic (CIS) from inflammatory lesions (Hunner) remains unclear. We aimed to synthesize existing evidence, assess real-world diagnostic performance, and identify key limitations and future directions.

A systematic review and meta-analysis were conducted in accordance with PRISMA guidelines. Studies from 2018 to 2024 were included if they reported validated AI-based cystoscopic diagnostic models for BC or IC/BPS. Risk of bias and heterogeneity were evaluated across studies. Five representative models were analyzed:
・Meta-Fusion RCNN (Lin et al.) – AI-based bladder cancer detection using white light imaging (WLI) and narrow-band imaging (NBI).
・RGB Deep Learning (Yoo et al.) –  Tumor grading and CIS detection using RGB image classification.
・AINAFHIC (Ueda et al.) – AI model for Hunner lesion detection in IC/BPS (presented at the 2022 Annual Meeting of the American Urological Association).
・Deep CNN (Eminaga et al.) – Classification of CIS and inflammatory bladder lesions.
・Deep Learning Model (Iwaki et al.) AI-assisted recognition of Hunner lesions in IC/BPS.
Pooled sensitivity, specificity, and accuracy were calculated using weighted averages. Diagnostic odds ratio (DOR) was also computed to evaluate overall performance in differentiating CIS from Hunner lesions. Risk of bias was assessed qualitatively, and heterogeneity was considered in pooled estimates.

AI-assisted cystoscopy demonstrated high diagnostic performance across studies:
・Pooled sensitivity: 91.8%
・Pooled specificity: 93.5%
・Overall accuracy: 92.6%
However, limitations emerged in CIS vs. Hunner lesion differentiation:
・Meta-Fusion RCNN (Lin et al.) improved bladder cancer detection, though it did not address CIS differentiation specifically.
・RGB-Based Model (Yoo et al.) achieved 97.2% sensitivity and 98.0% specificity for CIS; however, specificity dropped when 
    distinguishing CIS from Hunner lesions using NBI alone.
・AINAFHIC (Ueda et al.) reported 90.5% sensitivity in detecting Hunner lesions, though its specificity for ruling out CIS was limited.
・Deep CNN (Eminaga et al.) achieved 93.5% sensitivity and 94.0% specificity in differentiating CIS from inflammatory lesions.
・Deep Learning Model (Iwaki et al.) showed promising performance in identifying Hunner lesions, though visual overlap with CIS 
     posed diagnostic limitations.
The pooled diagnostic odds ratio (DOR) was 0.821, indicating modest discriminative ability in differentiating CIS from inflammatory lesions.

AI models offer strong diagnostic performance in bladder cancer detection and CIS identification. However, current models are limited in distinguishing CIS from Hunner lesions, particularly when relying on NBI. This highlights the need for further model refinement.
Multi-modal training with WLI, NBI, and photodynamic diagnosis (PDD), along with federated learning across institutions, may enhance lesion discrimination and generalizability. Real-time clinical integration remains essential for practical implementation.

This systematic review and meta-analysis confirm the promising diagnostic performance of AI-assisted cystoscopy. However, differentiating CIS from Hunner lesions remains a critical challenge due to their visual similarity. Future AI development should incorporate multi-modal imaging and collaborative learning frameworks to achieve reliable clinical application.

Lin J, Pan Y, Xu J, Bao Y, Zhuo H. A meta-fusion RCNN network for endoscopic visual bladder lesions. Computerized Medical Imaging and Graphics. 2022;3:29–35.
Yoo JW, Kim H, Park S, et al. Deep learning diagnostics for bladder tumor identification and grade prediction using RGB method. Scientific Reports. 2022;14:2003–2015.
Iwaki T, Yoshida T, Nakagawa M, et al. Deep learning models for cystoscopic recognition of Hunner lesion. European Urology Open Science. 2023;15:180–192.