Ilaria Jansen - Automated grading of urothelial cell carcinoma of the bladder

Background & objectives

Although histological grade is an important predictor for recurrence and progression in nonmuscle invasive bladder cancer (NMIBC), the reproducibility is low. The aim of this study is to investigate the potential value of a deep learning architecture for the grading of urothelial cell carcinoma (UCC) of the bladder using histology slides of transurethral resection of bladder tumor (TURBT) specimen and assess its accuracy by the comparison with the consensus grading of three pathologists.


Histological glass slides of patients with NMIBC who underwent a TURBT between 2000- 2016 in three hospitals in the Netherlands were included. The slides were independently reviewed by three pathologists, assigning the WHO'73 and WHO'04 grade, resulting in a four-tiered grading scheme. The slides were digitized, manually annotated by an expert observer and subsequently checked by a uropathologist. Firstly, a U-Net was trained to segment urothelium. Based on these segmentations, a ImageNet pre-trained DenseNet was trained for automated grading of the urothelial lesions based on the consensus score of the three pathologists.


In total, 328 tissue samples of 232 patients were included. In 93% of the slides, the urothelium was accurately detected. In another 21% false positive regions were detected. Incorrect classification was mainly in slides with extensive color loss or in regions with inflammation. Detailed data on the UCC grading will be presented at the conference.


This feasibility study demonstrates the potential value of deep learning methods in classification of urothelial lesions.