A hybrid model for diabetic retinopathy and diabetic macular edema severity grade classification

作     者：Sabeena, A. S. Jeyakumar, M. K. 

作者机构：Noorul Islam Ctr Higher Educ Dept Comp Sci & Engn Kanyakumari 629180 Tamil Nadu India Noorul Islam Ctr Higher Educ Dept Comp Applicat Kanyakumari 629180 Tamil Nadu India 

出 版 物：《INTERNATIONAL JOURNAL OF DIABETES IN DEVELOPING COUNTRIES》 (Int. J. Diabetes Dev. Countries)

年 卷 期：2025年

页      面：1-15页

核心收录：

主　　题：Diabetic retinopathy Diabetic macular edema Segmentation Severity grade classification Optimization algorithm Deep learning 

摘      要：BackgroundDiabetic retinopathy (DR) is a common reason for permanent vision loss, particularly among the elderly population worldwide;diabetic macular edema (DME) is characterized by the accumulation of fluid or swelling in the macula, which can occur in any phase of DR progression. New studies have deepened our knowledge of the needs in eye care to improve methods for recognizing, addressing, diagnosing, and treating retinal *** work intends to present a scheme for disease severity classification following the processes: pre-processing, segmentation, and disease severity classification. Firstly, pre-processing is done by employing a median filter, followed by GD-STFA-based SwinUNet for segmentation of lesions into soft and hard exudates, hemorrhages, and ***, DR and DME severity classification is employed by AlexNet-DQN, which is proposed by combining AlexNet and DQN models. Moreover, the proposed AlexNet-DQN is trained using the Exponential Gradient Descent-Sea Turtle Foraging Algorithm (ExpGD-STFA), which is developed by including the exponential weighted moving average (EWMA) concept in *** presented mechanism attained better results when compared to traditional *** DR severity classification, the proposed mechanism attained an accuracy of 98.3%, sensitivity of 98.9%, specificity of 97.8%, and precision of 97.6%. Meanwhile, for DME severity classification, the presented mechanism attained an accuracy of 98.1%, sensitivity of 98.8%, specificity of 97.5%, and precision of 97.3%.