Performance of an artificial intelligence algorithm for reporting urine cytopathology.

Area of Science:

• Urology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Automated screening systems for urine cytology are lacking, unlike Papanicolaou tests.
• Urine cytology faces challenges with interobserver agreement and definitive diagnoses, even with The Paris System.
• Development of an image algorithm using computational methods for digitized urine cytology slides is reported.

Purpose of the Study:

• To develop and validate a computational algorithm for analyzing digitized urine cytology slides.
• To improve the accuracy and consistency of urine cytology diagnoses.
• To provide computer-assisted interpretation for urine cytology cases.

Main Methods:

• Digitization of 2405 archival ThinPrep urine cytology slides.
• Development of a deep learning pipeline with convolutional neural networks for whole slide image (WSI) analysis.
• Validation of the algorithm using a separate test dataset of routine clinical cases.

Main Results:

• Analysis of 1.9 million urothelial cells across 2405 WSIs.
• Algorithm achieved an area under the curve of 0.88 for predicting diagnoses.
• For high-grade urothelial carcinoma, sensitivity was 79.5% and specificity was 84.5%.

Conclusions:

• A computational algorithm for accurate analysis of urine cytology WSIs was successfully developed.
• The algorithm utilizes whole slide-level features and a cell gallery for user review.
• This provides computer-assisted interpretation for urine cytology, analogous to Pap test screening technology.