Deep Learning Based DR Medical Image Classification

Abstract

AbstractDiabetic retinopathy (DR) is a prevailing disease that causes blindness among diabetic patients. The timely intervention with the regular fundus photography screening is the efficient way to cope with this disease. Screening of a large number of diabetic patients urges to the computer-aided and fully automatic DR diagnosis. Deep neural networks are gaining more attention due to their effectiveness in various tasks. The diagnosis of the can be made automatic and accurate suggestions can be provided to DR patients. The classification of DR images is challenging and important step. Therefore, in this paper, we have proposed a learning-based DR image reconstruction followed by a classification approach for DR image classification. Initially, the learning-based image reconstruction approach is proposed with a multi-encoder decoder network and residual inception block to delve into the features of input medical image and convert it into a set of abstract features followed by the reconstruction of the input medical image. The features from encoder are the abstract version of actual image representation which can be used for robust reconstruction for the input image. Therefore, these encoded features are used for DR image classification. The results analysis of the proposed framework with existing SOTA feature extraction algorithms is conducted on the MESSIDOR database for DR image classification. From the results’ analysis, it is evident that the proposed reconstruction-based classification framework outperforms the existing SOTA feature extraction algorithms.KeywordsImage reconstructionResidual inception blockDiabetic retinopathyClassification