Transfer Learning-Based Classification of Gastrointestinal Polyps

Abstract

We used a deep learning model, called Inception V3, to classify colorectal polyps into: hyperplastic, serrated and adenoma lesions using colonoscopy images. Inception V3 is a convolution neural network (CNN) pre-trained on an extremely large dataset, which is based on multi-branch convolutional networks. Because we have a relative small dataset, we use transfer learning (TL) to transfer the optimal weights of hundreds of hours of training across multiple high-power GPUs. A dataset 152 instances containing 76 polyps belonging to the three lesion types was used. We re-trained the last five layers of Inception V3 with two-thirds of the images in the dataset. The results obtained with our new neural network model are satisfactory compared to other works and human experts.