Deep Learning Image Analysis Algorithms to Improve Oral Cancer Risk Assessment for Oral Potentially Malignant Disorders

Abstract Oral potentially malignant disorders (OPMD) are a group of mucosal diseases in the oral cavity with a risk of progressing to oral squamous cell carcinoma. Risk assessment is traditionally done through a combination of clinical and histologic evaluation. Leukoplakia is a common type of OPMD that is given a histologic grading score that is supposed to be related to its risk of progression. However, there is tremendous intra- and inter- observer heterogeneity in dysplasia grading, leading to variability and uncertainty in risk assessment and treatment planning. This also hinders the ability to study the biology of these lesions. We propose to use whole slide imaging on routine hematoxylin and eosin (H&E) stained sections in combination with deep learning methods to build a consistent risk scoring system for OPMD. Our methods will identify cell, nucleus, and tissue architectural features relevant to risk of progression in OPMD. These features will be tested in a large retrospective case-control study and then validated prospectively. We will also explore combining them with genomic and immune biomarkers in order to improve the prognostic power and explore the biolo gy of progression in OPMD. We hope that these efforts will improve and standardize risk assessment for OPMD. This could lead to improved treatment and prevention options by enabling risk stratification and allowing future clinical trials be conducted in a more uniform patient cohort. Similarly, it could improve our understanding for the biology of OPMD and the process of progression to cancer.