Developing a Machine Learning Algorithm for Identifying Abnormal Urothelial Cells: A Feasibility Study.

INTRODUCTION

Urine cytology plays an important role in diagnosing urothelial carcinoma (UC). However, urine cytology interpretation is subjective and difficult. Morphogo (ALAB, Boston, MA, USA), equipped with automatic acquisition and scanning, optical focusing, and automatic classification with convolutional neural network has been developed for bone marrow aspirate smear analysis of hematopoietic diseases. The goal of this preliminary study was to determine the feasibility of developing a machine learning algorithm on Morphogo for identifying abnormal urothelial cells in urine cytology slides.

METHODS

Thirty-seven achieved abnormal urine cytology slides from cases with the diagnosis of atypical urothelial cells and above (suspicions or positive for UC) were obtained from 1 hospital. A pathologist (J.R.) reviewed the slides and manually selected and annotated representative cells to feed into Morphogo with following categories: benign (urothelial cells, squamous cells, degenerated cells, and inflammatory cells), atypical cells, and suspicious cells. Initial validation of the algorithm was performed on a subset of the original 37 cases. Urine samples from additional 12 unknown cases with various histological diagnoses (6 cases of high-grade urothelial carcinoma (HGUC), 1 case of low-grade urothelial carcinoma (LGUC), 1 case of prostate adenocarcinoma, 1 case of renal cell carcinoma, and 4 cases of non-neoplastic conditions) were collected from another hospital for initial blind testing.

RESULTS

A total of 1,910 benign and 1,978 abnormal (atypical and suspicious) cells from 37 slides were annotated for developing and training of the algorithm. This algorithm was validated on 27 slides that resulted in identification of at least 1 abnormal cell per slide, with a total of 200 abnormal cells, and an average of 7.4 cells per slide. Of the 12 unknown cases tested, the original cytology was positive for tumor cells in 2 HGUC samples. Morphogo was abnormal (atypical or suspicious) for 6 samples from patients with UC, including one with LGUC and one with prostate adenocarcinoma.

CONCLUSION

Morphogo machine learning algorithm is capable of identifying abnormal urothelial cells. Further validation studies with a larger number of urine samples will be needed to determine if it can be used to assist the cytological diagnosis of UC.