Qualoscopy: convolutional network for colonic polyp classification in the browser

Author: Andrew Ninh

Colonoscopy with polypectomy is the gold standard screening procedure for colorectal cancer prevention. Though this procedure is generally performed in adults, there are similar applications in pediatrics -- particularly with children or adolescents with polyposis syndromes such as familial adenomatous polyposis (FAP) and juvenile polyposis syndrome. When performing a colonoscopy or sigmoidoscopy on children or adolescents, pediatric gastroenterologists may find it useful to be assisted by an AI that helps identify polyps in real-time and classify polyps based on morphological characteristics and meta-feature vectors (demographics, reason for procedure, location of polyp, etc.). We propose a deep learning approach to identifying and classifying polyps in the colon. The initial model will be a convolutional neural network (ConvNet) trained on 4,000 images of different segments of the adult colon with polyps and another 4,000 images without polyps. As more data is collected in real time, we will identify the model's weaknesses and edge cases and continue to re-train and refine the algorithm. It would be interesting to measure the model's efficacy on pediatric populations. We hypothesize the model will perform with a similar efficacy in pediatric populations which will be determined by measuring the model's precision and recall and F1 score when tested on pediatric colon image data. The model will be implemented in an HTML web application (Qualoscopy) in the browser. It will initially be implemented with the image-capture button to classify images as they're taken. The eventual goal is to implement the algorithm as a real-time filter for the endoscope video stream which can either be done using JavaScript in the front-end or using FFMPEG server-side.

Co Author/Co-Investigator Names/Professional Title: William Karnes, MD; Associate Clinical Professor at UCI Pierre Baldi, PhD; Chancellor's Professor at UCI