RTSO-DMQLN: REMORA TUNA SWARM OPTIMIZATION BASED DEEP MAXOUT Q LEARNING NETWORK FOR DIABETIC RETINOPATHY DETECTION

Diabetic Retinopathy (DR) is one of the important reasons for vision disability, mostly among working-age people worldwide. Early identification of DR significantly reduces the risk of vision loss. Although various automated DR detection approaches have been proposed, many of them have failed to provide a high-accuracy rate in real-world scenarios. To subdue these difficulties, a novel Remora Tuna Swarm Optimization-based Deep Maxout Q Learning Network (RTSO-DMQLN) is proposed for enhanced DR detection. The method starts with the pre-processing of input images using a Nonlocal Means filter (NLM). After that, [Formula: see text]-Net is used for lesion segmentation and Dense [Formula: see text]-Net is used for blood vessel segmentation. The features from the original images and segmented images are extracted in the next step. These extracted features are used for the classification through the RTSO-DMQLN model. RTSO-DMQLN is an integration of Deep Maxout Network (DMN) and Deep Q Learning (DQN) as well as Remora Tuna Swarm Optimization (RTSO). The outcomes of DR are classified into five categories: normal, mild, moderate, severe Nonproliferative DR (NPDR), and Proliferative DR. Experimental results show that the proposed system achieved an accuracy of 92.2%, True Positive Rate (TPR) of 92.8%, True Negative Rate (TNR) of 93.7%, Positive Predictive Value (PPV) of 90.7% and Negative Predictive Value (NPV) of 91.8%. Compared with recent DR classification models such as MobileNet, Active Deep Learning-convolution Neural Networks (ADL-CNN), Deep Neural Network (DNN) with Chronological Tunicate Swarm Algorithm (CTSA), DNN, and Deep Residual Network (DRN)– Remora Tuna Swarm Optimization (RTSO), the proposed model achieved an improvement of 12.690%, 9.436%, 7.809%, 6.399%, and 3.362% in accuracy.