Transfer Learning for Colonic Polyp Classification Using Off-the-Shelf CNN Features

Recently, a great development in image recognition has been achieved, especially by the availability of large and annotated databases and the application of Deep Learning on these data. Convolutional Neural Networks (CNN's) can be used to enable the extraction of highly representative features among the network layers filtering, selecting and using these features in the last fully connected layers for pattern classification. However, CNN training for automatic medical image classification still provides a challenge due to the lack of large and publicly available annotated databases. In this work, we evaluate and analyze the use of CNN's as a general feature descriptor doing transfer learning to generate off-the-shelf CNN's features for the colonic polyp classification task. The good results obtained by off-the-shelf CNN's features in many different databases suggest that features learned from CNN with natural images can be highly relevant for colonic polyp classification.