Colorectal image classification with transfer learning and auto-adaptive artificial intelligence platform

Abstract

In automatic (computer-based) interpretation of medical images, the use of deep learning techniques is limited because of the lack of large publicly available datasets. With just hundreds of samples (images) in a dataset, the application of deep learning techniques is very hard, and the results are under expectations. Training a multi layer convolutional neural network requires thousands or even millions of images for an acceptable level of correct classification. In this paper we will present a novel approach that can be used to solve computer vision related problems (e.g. medical image processing) even when only a small dataset of images are available for training. We will show that combining Transfer Learning and some auto-adaptive artificial intelligence algorithms we can obtain very good classification rates even with the use of a limited dataset. As a demonstration of the effectiveness of our approach we will show the use of this technique to solve the polyp detection problem in endoscopic image sets. We show that using just a subset of the available images (from the original dataset containing 4000 images) the results are comparable with the case when all the images were used.