Jointly Subspace Hashing for Medical Image Retrieval

Abstract

The explosive growth of medical images brings about massive amounts of high-dimensional images. The medical image retrieval technique is effective in selecting some similar medical images for medical analysis. A vast majority of existing medical image retrieval methods aim to extract a single feature to represent a medical image. Although these methods have improved the performance, they ignore that different characteristics express different information and they are an indispensable part of content information. Moreover, conventional features consume much storage, which is not conducive to data storage. Therefore, to reduce the loss of information and save storage space, we propose a novel jointly subspace hashing (JSSH) method for medical image retrieval. We solve the high-dimensionality of medical by the simplest image segmentation technology, and then study different subspace projection matrices. Finally, we integrate these matrices into the hash learning model, and build an objective function and learn a series of discriminative binary codes. By conducting comprehensive experiments on three medical image benchmark datasets, we demonstrate the effectiveness of our proposed JSSH.