Keywords: Computer Aided Diagnosis, Siamese Network, Polyp Classification
TL;DR: Polyp classification using Siamese neural networks and k-NN.
Abstract: Colorectal cancer is one of the leading cause of cancer related deaths with increasing prevalence. One key factor in the likelihood of adenomatous cell differentiation is polyp diameter. There exist a significant cut-off value of 10 mm which clinicians use in diagnosis management. We propose a novel method to classify polyp size into above or below 10 mm classes based on a Siamese network. In a first step, a Siamese networks is trained to build a high dimensional feature embedding extracted for each polyp size. In as second step, we use a k-NN approach to classify polyp sizes based on the distance between the feature embedding of the input image, and the whole embedding space learned by the Siamese network. This method allows for better binary classification of the sub- and sup- 10 mm polyp size classes. Our data consist of around 55,000 images from 129 movies classified into various polyp sizes ranging from 1-15 mm. We trained our model on 10,746 images, and tested on 2,688 images equally split into each binary category. We obtained 79.2% in feature classification and 95.7% in polyp size classification.
Code Of Conduct: I have read and accept the code of conduct.
Remove If Rejected: Remove submission from public view if paper is rejected.