AI-Driven Comparative Study of High-Grade Cell Features in Urine Cytology with Biopsy Correlation

Background:Urine cytology is a common and cost-effective screening and diagnostic tool for detecting high-grade urothelial carcinoma (HGUC). The Paris System (TPS) for reporting urine cytology defines the cytomorphologic characteristics for the reliable diagnosis of HGUC and emphasizes the nuclear-cytoplasmic (N/C) ratio as a key criterion. Additionally, nuclear size is also a relevant reference for reporting urine cytology. Currently, there is a lack of tools for generating quantitative and numerical data on N/C ratio and nuclear size from large cell populations, hampering the correlation between TPS criteria, biopsy, and cytology findings. To address these subjective challenges, we developed AIxURO, an AI-powered software designed to generate cell-level representations from a whole slide image. Based on TPS morphological features, N/C ratio, and nuclear size, AIxURO performs quantitative analysis to assist in urine cytology reporting. This study also aimed to explore the cytologic-histopathologic correlations in biopsy-confirmed carcinoma in situ (CIS) and HGUC using AIxURO.Methods: A multi-institutional retrospective study (4 hospitals) submitted 242 urine cytology slides (74 AUC, 56 SHGUC, and 112 HGUC) and their corresponding positive biopsy specimens collected within six months. Among these, 212 biopsies were diagnosed as high-grade urothelial carcinoma (Biopsy-HGUC) and 30 as carcinoma in situ (Biopsy-CIS). The 242 corresponding abnormal urine cytology slides were prepared using three preparation methods: Cytospin (162 slides), ThinPrep UroCyte (60 slides), and BD CytoRich (SurePath, 20 slides), followed by Papanicolaou staining. The AIxURO software categorized cells as Atypical Urothelial Cells (atypical cells) or Suspicious for High-Grade Urothelial Carcinoma (suspicious cells) and presented the top 24 'suspicious' cell images considered to have the highest risk of malignancy in a gallery for review. Statistical significance was evaluated using Kruskal-Wallis tests.Results:Among 242 cytology slides, the software identified a total of 124,980 abnormal cells (16,662 suspicious cells and 108,318 atypical cells). The N/C ratio and nuclear size of each cell were analyzed. The average N/C ratio of the top 24 AI-selected suspicious cells (0.66, 95% CI: 0.65-0.66) was larger than all suspicious cells detected (0.65, 95% CI: 0.64-0.65, p = 0.0035) and significantly larger by 15.8% than categorized atypical cells (0.57, 95% CI: 0.57-0.57, p <0.0001), respectively. The average nuclear size of the top 24 suspicious cells (110.5 &micro;m&sup2;, 95% CI: 105.7-115.3 ) was significantly larger than total atypical cells (85.8 &micro;m&sup2;, 95% CI: 83.1-88.5, p <0.0001).In the cytology HGUC group, the N/C ratio of the top 24 suspicious cells (0.66, 95% CI: 0.65-0.67) was larger than the total number of suspicious cells (0.65, 95% CI: 0.64-0.65, p = 0.0056) and significantly larger than total atypical cells (0.57, 95% CI: 0.57-0.57, p <0.0001). The N/C ratio of the top 24 suspicious cells was also larger than the total atypical cells, with a statistically significant difference (p < 0.0001) in cytology AUC+ and SHGUC+ groups. Additionally, in the HGUC group, the nuclear size of the top 24 suspicious cells (108.1 &micro;m2) was larger than those of atypical cells (84.5 &micro;m2, p <0.0001). Similar results were also found in cytology AUC + and SHGUC+ groups. No significant difference in average nuclear size was found among the three sample preparation types (Cytospin, UroCyte, and CytoRich) for either suspicious or atypical cells.Conclusions:The AIxURO software using Whole Slide Imaging demonstrates a significant advancement in quantitative cytology analysis, offering a rapid, consistent, and precise evaluation of the nuclear size and N/C ratio. In this study, significantly larger N/C ratios and nuclear sizes occurred in the top 24 suspicious cells displayed in the gallery than in the total suspicious cells, or total atypical cells analyzed from urine cytology slides correlated with 242 positive biopsy specimens.These preliminary findings suggest that the cell characteristics and classification of abnormal urine cytology cells based on risk stratification with gallery presentation by AIxURO provide a valuable tool for quantitative analysis to assist in decision-making for urothelial carcinoma detection. The ability to rapidly and consistently quantify cytologic characteristics has the potential to mitigate interobserver discrepancies and enhance patient care.

Learning Objectives

1. AIxURO identified and analyzed 124,980 abnormal cells from 242 urine cytology slides (AUC, SHGUC, and HGUC), distinguishing between suspicious and atypical cells with high precision in a multi-institutional study.  
2. Significantly larger N/C ratios and nuclear sizes occurred in the top 24 suspicious cells displayed in the gallery than in the total suspicious cells, or total atypical cells analyzed from urine cytology slides correlated with 242 positive biopsy specimens.    
3. This preliminary study demonstrated that AIxURO could effectively provide a valuable tool for quantitative analysis of suspicious cells and improve diagnostic decision-making in urothelial carcinoma detection.