Adenoma Detection Through Deep Learning: 2017 Presidential Poster Award

Abstract

Introduction: For every percentage increase in adenoma detection rate (ADR), risk of interval colorectal cancer is reduced by 3%. Adenoma prevalence is estimated to be 50%. Ideally, ADR should reflect adenoma prevalence, yet ADR varies between 5 and 55% among colonoscopist. We hypothesize that computer-assisted polyp identification could help bring low ADRs closer to true prevalence and reduce interval colorectal cancer incidence. Methods: Utilizing our colonoscopy quality database (Qulaoscopy, Docbot, Inc,), we applied the VGG-16 model as starting point (see [1]), which is a convolutional neural network (CNN) with 16 convolutional layers that was trained on natural images as part of the ImageNet challenge. We then “fine-tuned” this CNN on a set of 9000 screening colonoscopy images originating from all locations in the colorectum (and half of which contain polyps) to classify the presence or absence of polyps in images. The RGB images are scaled to a size of 224x224 pixels for this purpose and the CNN is trained using ten-fold crossvalidation, i.e. we repeat the training and testing ten times with different parts of the data for training/testing the model. Results: On the task of classifying polyp versus non-polyp containing images we obtain an accuracy of 96% and AUC (area under the ROC curve) of 0.99, which is an almost perfect score (the maximum attainable AUC is 1). At a processing rate of 170 images per second, the algorithm was easily applied to live video. Conclusion: Our deep learning methology achieved extremely high accuracy for polyp detection in images and is proven applicable to live video. We predict that application of this technology during colonoscopy will assist all colonoscopist to achieve ADRs that approach true adenoma prevalence.Figure: AUC (area under the ROC curve) for Polyp Detection.