V12-01 REAL-TIME AUGMENTED BLADDER TUMOR DETECTION WITH DEEP LEARNING

Abstract

INTRODUCTION AND OBJECTIVE: White light cystoscopy is the standard imaging modality for bladder cancer detection, treatment and surveillance, but around 15% of tumors are missed. Enhanced cystoscopy technologies such as blue light cystoscopy improve tumor detection, but their adoption remains limited. Recently, we developed CystoNet, a deep learning algorithm capable of automated annotation of benign bladder landmarks and detection of bladder tumors without additional modification to standard cystoscopy instrumentation. Herein we describe real time integration of CystoNet in real world clinical workflow including office cystoscopy and transurethral resection. METHODS: The study received IRB approval. CystoNet, which previously was trained, validated and tested to achieve high accuracy in detecting bladder tumors, was integrated to cystoscopy systems in the clinic and in the operating room (both flexible and rigid) to achieve realtime detection of bladder tumors. There were no modifications made to the cystoscopy instruments. The cystoscopy output images were captured directly to a laptop equipped with GPU processors. The resulting CystoNet results were then displayed side-by-side the actual cystoscopy images in real-time. RESULTS: To date, 21 subjects have prospectively undergone cystoscopy with CystoNet. 8 cystoscopies were performed in the clinic, and 13 were performed in the operating room. No adverse events were noted. CystoNet was able to detect papillary and flat lesions in real-time. Benign features such as bubbles were also appropriately identified. Corresponding bladder tissue specimen were sent for histopathology to determine accuracy. Of the 12 tumors confirmed on pathology, all 12 were detected in real-time by CystoNet. CONCLUSIONS: We report the implementation of a deep learning system to cystoscopy in real-time with the goal of achieving improved detection of bladder tumor. This study thus far demonstrates high accuracy in detecting bladder tumors in real-time. The technology can be applied to various imaging modalities including blue light cystoscopy without significant modifications or administration of medications. This augmented cystoscopy may assist urologists in achieving improved cancer detection.