Artificial Intelligence‐Assisted Urine Cytology for Noninvasive Detection of Muscle‐Invasive Urothelial Carcinoma: A Multi‐Center Diagnostic Study with Prospective Validation

Abstract Accurate preoperative diagnosis of muscle invasion (MI) is critical for urothelial carcinoma (UC) management. The aim is to evaluate whether artificial intelligence (AI) model based on urine cytology can accurately detect MIUC and compare its performance with radiologist assessments. UC patients underwent liquid‐based urine cytology from four centers are included for model development/validation. Performance of the precision urine cytology AI solution for MI (PUCAS‐M) is validated across multicenter cohorts and compared to radiologists’ assessments (including CT/MR, MR accounted for 40.7%). Clinical utility is assessed for initial diagnosis, recurrence detection, and neoadjuvant therapy. PUCAS‐M achieves an area under the receiver operation curve (AUROC) of 0.857 (95% CI: 0.820‐0.895) in the whole validation cohort, which is significantly higher ( P ‐value = 0.005) than radiologists (0.773, 95% CI: 0.727–0.818). The integration of radiologists' diagnosis and PUCAS‐M (mPUCAS‐M) significantly increases the sensitivity of radiologists from 63.9% to 83.3% in bladder cancer and from 76.9% to 90.3% in upper‐tract UC. Lastly, in the neoadjuvant therapy subgroups, mPUCAS‐M maintains an improved AUROC (ranging from 0.857‐0.865), whereas radiologist assessments' performance decline. PUCAS‐M provides accurate, non‐invasive MI detection method, particularly valuable for equivocal imaging. Integration with clinical data enhances diagnostic precision, offering a scalable solution for UC management.