Accurate Classification of Dysplasia in Inflammatory Bowel Disease Patients Using Deep Learning

Patients with inflammatory bowel disease (IBD) are at high risk for developing dysplasia and colorectal cancer. The early and accurate detection and treatment of dysplasia forms the main strategy to reduce mortality from colorectal cancer in IBD patients. Detecting such dysplasia is challenging because of the subtle, unconventional, multi-focal nature of the lesions. In this work, we develop an approach for accurate classification of dysplasia in IBD patients using Bayesian deep learning. We modify existing deep learning models to perform Bayesian approximation for achieving higher classification accuracy than a deterministic deep learning model. Specifically, we propose to insert one or more densely connected layers before the final densely connected layer of a model that performs classification. Each newly inserted layer is followed by a dropout layer. These inserted dropout layers are enabled during training and inference. Instead of obtaining a single prediction by a deterministic model for a given test input, we obtain a distribution of predictions and then compute the most probable prediction. We evaluated our approach using 60+ digital slides of histopathology tissue sections containing three different types of dysplasia in IBD patients. Our best Bayesian deep learning model achieved an accuracy of 97.37%, 93.23%, and 98.16%, respectively for the three dysplasia types using patch-wise classification.