Leveraging deep preference learning for indexing and retrieval of biomedical images

Abstract

This paper presents an original framework based on deep learning and preference learning to retrieve and characterize biomedical images for assisting physicians in diagnosing complex diseases with potentially only small differences between them. In particular, we use deep learning to extract the high-level and compact features for biomedical images. In contrast to the traditional biomedical algorithms or general image retrieval systems that only consider the use of pixel and/or hand-crafted features to represent images, we utilize deep neural networks for feature discovery of biomedical images. Moreover, in order to be able to index the similarly referenced images, we introduce preference learning in a novel way to learn what kinds of images we need so that we can obtain the similarity ranking list of biomedical images. We evaluate the performance of our system in detailed experiments over the well-known available OASIS-MRI database for whole brain neuroimaging as a benchmark and compare it with those of the traditional biomedical and general image retrieval approaches. Our proposed system exhibits an outstanding retrieval ability and efficiency for biomedical image applications.