(UroToday.com) The 2026 American Urological Association annual meeting featured a non-invasive bladder cancer session and a presentation by Dr. Masatomo Kaneko discussing a fully automated artificial intelligence system for urine cytology and results from a large multicenter external validation. Although numerous urinary biomarkers have emerged, urine cytology remains a key component of high-grade urothelial cancer (HGUC) surveillance. In this study, presented at the AUA 2026 annual meeting, a fully automated artificial intelligence system for urine cytology was validated on a large multicenter external dataset.
Urine cytology slides were collected from patients with suspicion of urothelial carcinoma at 6 institutions in Japan and the United States. Two board-certified cytotechnologists and cytopathologists independently reviewed and labeled the slides according to the Paris system. A deep learning artificial intelligence model was developed to automate analysis on digital cytology slides and report a cytology diagnosis based on the Paris system, highlighting the top 15 suspected high-grade urothelial cancer cells. In institutions using the double-slide smear method, the slide with the higher prediction was adopted. After the cytology, the presence or absence of high-grade urothelial cancer was histologically confirmed by surgery, or close follow-up over 1 year with no evidence of urothelial carcinoma as a negative control. The detection of high-grade urothelial cancer from cytology slides was evaluated by the receiver operating characteristic, and the accuracy and sensitivity of artificial intelligence were compared with those of pathologists by the McNemar test, at the cutoff where the artificial intelligence achieved the same specificity as the pathologists.
The artificial intelligence model was trained on 475 slides from two institutions and externally validated on 811 unseen slides from four other institutions. In the test cohort, 58% of patients had histologically confirmed high-grade urothelial cancer. Overall, the area under the receiver operating characteristic curve (AUC) of the artificial intelligence was 0.79. The performance of artificial intelligence versus pathologists was comparable:
- Accuracy: 67% versus 66% (p = 0.5)
- Sensitivity: 49% versus 47% (p = 0.5):
- Specificity: 93% versus 93% (matched)
Across institutions, the ranges of AUC, accuracy, sensitivity, and specificity for artificial intelligence were 0.67–0.81, 63–72%, 21–64%, and 62–98%, respectively, which were comparable to those of the pathologists (all p > 0.05):
Dr. Kaneko concluded this presentation discussing a fully automated artificial intelligence system for urine cytology with the following take-home points:
- A fully automated artificial intelligence system for evaluating urine cytology demonstrated robust and consistent performance, detecting histologically confirmed high-grade urothelial cancer with accuracy and sensitivity comparable to those of board-certified pathologists across large external institutions
- This artificial intelligence system may automate the pre-screening process and streamline the urine cytology workflow, thereby effectively assisting uropathologists
Presented by: Masatomo Kaneko, MD, PhD, Research Scholar, University of Southern California, Los Angeles, CA
Written by: Zachary Klaassen, MD, MSc – Urologic Oncologist, Associate Professor of Urology, Georgia Cancer Center, Wellstar MCG Health, @zklaassen_md on Twitter during the American Urological Association (AUA) 2026 Annual Meeting, Washington, DC, Fri, May 15 – Mon, May 18, 2026.