Colonic Polyp Classification with Convolutional Neural Networks

Texture patch classification is an important task in many different computer-aided medical systems. Convolutional Neural Networks (CNN's) have become state-of-the-art for many computer vision tasks in recent years. In this paper, we propose the use of CNN's for the automated classification of colonic mucosa for colon polyp staging in the context of colon cancer screening. This deep learning approach has the property of extracting features and classifying images in the same architecture by exploiting directly the input image pixels being successful in handling distortions such as different light conditions, presence of partial occlusions, etc. For this type of deep learning approach it is common to require that the database contains large amounts of data, which is quite rare in the medical field. The method proposed allows the use of small patches (subimages) to increase the size of the database as well to classify different regions in the same image. We show experimentally that this model is more efficient than some of the commonly used features for colonic polyp classification.