A bi-directional Long Short-Term Memory-based Diabetic Retinopathy detection model using retinal fundus images

Abstract

Images of the retina are widely used for diagnosing fundus disease. Low-quality retinal photos make it hard for computer-aided diagnosis systems and ophthalmologists to make a clinical diagnosis. In ophthalmology, precision medicine is based partly on the quality of retinal images. Diabetic Retinopathy (DR) is a common complication of diabetes mellitus that causes iris damage. It is difficult to detect and, if not detected early, can result in blindness. Convolutional neural networks are gaining popularity as an effective deep learning (DL) approach for medical image analysis. This study suggests using deep learning approaches at various stages of the fundus image-based diagnostic pipeline for diabetic retinopathy (DR). Many fields, including medical image classification, have adopted DL representations. Using retinal fundus images, we propose a bi-directional extended short-term memory-based diabetic retinopathy detection model. By examining images of the retinal fundus, the Bi-directional Long Short-Term Memory (LSTM) method can detect and classify various grades of DR. As a preprocessing step, the proposed model uses the Multiscale Retinex with Chromaticity Preservation (MSRCP) method to increase the difference of fundus pictures and progress the short difference of medicinal views. To prepare satisfactory results for image processing, multiscale retinex with chromaticity preservation is used. However, choosing the parameters’ values, such as the Gaussian scales, gain, offset, etc., is the main difficulty with the retinex algorithm. To achieve a practical effect, these parameters must be adjusted. The main goal of the suggested method is to obtain the ideal values for the parameters used in the MSRCP algorithm. Also, photos that have already been processed are used to make feature vectors with the help of an efficient net-based feature extractor that uses deep learning. Many experiments use the benchmark Methods to Evaluate Segmentation and Indexing Techniques in the Field of Retinal Ophthalmology (MESSIDOR) dataset. The results are analyzed in terms of various evaluation factors. The results show that the Bi-LSTM-MSRCP technique is better at diagnosing DR than more modern methods.