Too good to be true? AI performance in CT-based urinary stone detection and classification: EAU endourology systematic review and diagnostic accuracy meta-analysis


Talyshinskii A., Tzelves L., Gauhar V., Nedbal C., Gorgotsky I., Gadzhiev N., ...Daha Fazla

World Journal of Urology, cilt.44, sa.1, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Derleme
  • Cilt numarası: 44 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1007/s00345-026-06885-1
  • Dergi Adı: World Journal of Urology
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, EMBASE, Gender Studies Database, MEDLINE, Academic Search Ultimate (EBSCO), Biomedical Reference Collection: Corporate Edition (EBSCO), Health Research Premium Collection (ProQuest), Pharma Collection (ProQuest)
  • Anahtar Kelimeler: Artificial intelligence, Classification, Computed tomography, Detection, Urolithiasis
  • İstanbul Üniversitesi Adresli: Evet

Özet

Purpose: To evaluate pooled performance of artificial intelligence (AI) in computed tomography (CT)–based urinary stone detection and classification and identify methodological factors influencing results. Methods: This review was registered in PROSPERO (CRD420261319940). A systematic search of PubMed, Cochrane Library, and Google Scholar was conducted up to March 15, 2026. Studies evaluating AI-based detection or classification of urolithiasis on CT with available validation/test cohort data (sensitivity/specificity or 2 × 2 tables) were included. A bivariate random-effects model generated pooled sensitivity, specificity, and hierarchical summary receiver operating characteristic (HSROC) curves. Meta-regression explored moderators related to dataset composition and AI pipelines. Risk of bias was assessed using QUADAS-AI. Results: 14 studies (five detection-based, nine classification-based) were included. Overall pooled sensitivity and specificity were 0.967 and 0.979 (AUC 0.987). Classification studies demonstrated higher pooled sensitivity (0.977) and specificity (0.990) compared with detection studies (0.935 and 0.919; AUC 0.945). Significant heterogeneity was observed and influenced by training sample size, class imbalance, presence of bladder stones or other pathologies, and architectural choices. Most studies had high risk of bias, lacked external validation, and frequently used slice-level or pre-selected datasets, potentially overestimating performance. Conclusion: Substantial methodological heterogeneity and high risk of bias limit clinical generalizability. Detection studies better reflect real-world workflows and show lower performance. Current evidence does not support routine clinical implementation without standardized reporting and robust external validation.