A Bladder Cancer Grading System Using Deep Neural Network Architectures

Histopathologists grade bladder cancer by looking at histology images or glass-slides of bladder tissues. Grading is an important stage of bladder cancer diagnosis and key requirement to determine the proper course of treatment. This work presents a machine-learning-based platform, known as Bladder Cancer Grading (BCG) system, a platform designed and developed based on deep neural network architectures. The main input to BCG is a high resolution image of bladder tissue scanned from glass slide, known as Whole Slide Image (WSI). In learning stage, BCG breaks a set of sample WSI images into equally-sized square tiles in order to build its learning dataset. In projection stage, similar to learning stage, BCG breaks down any given input WSI image into tiles and projects every single tile using the selected trained model. BCG uses a 4-tier grading scheme to grade each tile. The grading scheme is derived from World Health Organization (WHO) 1973/2004 bladder-cancer-grading-schemes [1] that are globally accepted and practiced by many specialists including pathologists working in University Health Network(UHN) in Toronto General 1 and Mount Sinai Hospitals 2. Using its distinct tiling design approach, BCG has introduced a grading scheme that projects a decimal grading value per each slide unlike existing practice that assigns a discrete grade value between 1 and 4 to each slide. The team of senior pathologists who assisted labeling BCG training dataset believe this new grading approach provides a better state of progression in bladder cancer resulting in more precise diagnosis and better treatment procedures. The choice of learning model(s) in BCG is configurable. Any deep architecture model could be plugged in, trained, and used by BCG. Some trained models developed by this platform have shown promising grading accuracy (more than 97% in verification/testing, above 85% in real projection, and 97+ specificity rate). BCG has also shown a highly consistent intro-observatory results. The combination of a loosely coupled architecture and fully integrated utilization of tiles in all stages of its execution have made BCG a universal, expandable, scalable, and versatile platform that could be configured and deployed in distributed running environments.