Abstract B13: Applying machine learning for urine cytology—computational urothelial carcinoma analysis and diagnosis

Abstract

Abstract Urothelial carcinoma is estimated to be the fourth most common cancer among men in the U.S. in 2019. The key to successful management of this cancer is early detection and timely monitoring after treatment. Urine cytology is an inexpensive, noninvasive, and easily available diagnostic method of urine analysis, but the usage of this tool for urothelial malignancy is still not satisfactory in its sensitivity and specificity. It is time-consuming and labor-intensive to find rare neoplastic cells from hundreds or even thousands of a variety of types of cells in the background, even by well-trained cytopathologists or cytotechnicians. Therefore, it is a crucial issue to avoid false-positive and false-negative results when applying urine cytology in clinical practice. Recent advances in artificial intelligence (AI) offer potential solutions to these issues through the development of an automated, impartial, and specialized computation model to 1) reduce false-negative rates and interpretation times and 2) assist pathologists with reporting using The Paris System for Reporting Urinary Cytopathology. In this study, we tested a deep learning-based image analysis model for cell classification and enumeration in urine cytology. Deidentified whole-slide images were digitalized after performing standard urine cytology examination. We used an “active learning” approach that requires minimal upfront training and improves with more usage. We collected 3 sub-images (total 3,335 cells) from one whole-slide image annotated by 3 domain experts to train the initial model. Cells were classified into seven categories: high-grade urothelial carcinoma (HGUC), cluster HGUC, atypical neoplastic cell, atypical reactive cell, inflammatory cell, epithelial cell, and unidentified cell. Domain experts reviewed and provided feedback to initiate the model training process. After the initial model was established, we performed a pilot study using 2 sub-images from five digitalized sample slides (total 10 sub-images) to further improve the training model. Our results showed that the AI model was able to learn the morphologies of six cell types. It was then able to automatically quantify total cell counts in each class that correlated with annotations by domain experts. We are now attempting to integrate this AI model with The Paris System to create a clinical decision support system for urologists, as there is a clear unmet clinical need to establish a computational model since The Paris System requires a significant amount of complex and accurate analysis, which many clinical decisions are based upon. In conclusion, our AI model demonstrates its value in providing an automated and unbiased method to analyze urine cytology with high throughput, accuracy, and reproducibility. Introduction and integration of our AI model with The Paris System will be a giant stride to optimize clinical practices in the field of urology in the imminent future. Citation Format: Wei-Lei Yang, Ching-Ming Lee, Mei-Ling Wu, Yu-Ching Peng, Ten-Jen Liu, Yi-Hsin Liu. Applying machine learning for urine cytology—computational urothelial carcinoma analysis and diagnosis [abstract]. In: Proceedings of the AACR Special Conference on Advances in Liquid Biopsies; Jan 13-16, 2020; Miami, FL. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(11_Suppl):Abstract nr B13.