Abstract
The Paris System (TPS) for Reporting Urinary Cytology is an international system with standardized terminology and cytomorphologic criteria that aims for better risk stratification to identify high-grade urothelial carcinoma in urinary cytology specimens. Based on TPS, the nuclear/cytoplasmic ratio (N/C ratio) of the observed urothelial cells is a crucial criterion and the cutoff values of 0.5 and 0.7 are used for the diagnosis of Atypical Urothelial Cells (AUC) and High-Grade Urothelial Carcinoma (HGUC)/Suspicious for HGUC respectively. Although the N/C ratio is a clear and measurable criterion, there are still challenges for adopting TPS into clinical practice. As urine cytology mainly relies on manual microscopic examination by cytopathologists and cytotechnologists, determining the N/C ratio through the naked eyes results in intra- and inter-observer variations and therefore generates subjective, inconsistent, and unreproducible reports. To overcome the aforementioned issues, the authors developed a deep learning-based image analysis model to assist calculating the N/C ratio and classifying urothelial cells computationally. Deidentified urine cytology slides were digitalized as whole-slide images (WSIs) for analyzing. An “active learning” method was used to improve results over time with only a small amount of upfront training data required. First, a small amount of annotated WSIs were provided to train the initial computational model. Next, the experts reviewed the results predicted by the initial model, and provide feedback for the model to learn and improve. The above steps were repeated by experts reviewing the “improved model” and providing feedbacks, thus improving the initial/intermediate models continuously until a satisfactory result was achieved. Preliminary results showed that our model was able to determine the areas of cytoplasm and nucleus in urothelial cells as well as calculate the N/C ratios using the WSI. Moreover, it enumerates the AUC and HGUC cells on the WSI and provides quantification and distribution information on the whole slide scale. This computational model provides a significant advantage as an objective tool to assist urine cytology analysis of the N/C ratio with greater accuracy and reproducibility.
Citation Format: Wei-Lei Yang, Chi-Bin Li, Yen-Chuan Ou, Yi-Sheng Lin, Tang-Yi Tsao, Ming-Chen Chang, Jen-Fan Hang, Tien-Jen Liu. A deep learning model assists urine cytology reporting with computational estimates of the nuclear/cytoplasmic ratios of the urothelial cells based on the Paris System [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-069.