Deep Supervised Hashing through Ensemble CNN Feature Extraction and Low-Rank Matrix Factorization for Retinal Image Retrieval of Diabetic Retinopathy

Abstract

During the past few decades, content-based image retrieval (CBIR) has been a prominent research area in medical image analysis. It enables retrieving images from an image database that are similar to a given query image. Numerous types of medical image retrieval approaches have been proposed by different research groups. In particular, supervised, deep neural network-based methods have achieved higher accuracy than others. However, they are computationally very expensive and an effective and comprehensive deep neural network-based retinal image retrieval model for diabetic retinopathy (DR) is not available in the literature. The principal objective of CBIR for DR is to efficiently retrieve retinal images that are semantically similar to a given query for effective treatment based on the severity stage of the disease. We propose to use a deep, supervised hashing approach in order to perform efficient retinal image retrieval, where we implicitly learn a good image representation along with a similarity-preserving compact binary hash code for each image by extracting features using an ensemble of deep convolutional neural networks through transfer learning and then feed these extracted features to an ANN classifier. This approach maps the image pixels to a lower-dimensional space and then generates compact binary codes to speedup the retrieval process. Moreover, our approach requires less memory and computational time, which can constructively accelerate the training process. Our experimental results show a considerable improvement compare to the other several state-of-the-art hashing techniques on the retinal dataset. We further analyze the effectiveness and efficiency of our approach using another medical dataset, KVASIR, which includes Gastrointestinal tract endoscopic imagery.