Abstract
Retinal detachment is a severely curable eye condition that becomes a genuine factor for the increased visual acuity worldwide. If neglected, it may result significant visual impairment in individuals aged 60 to 69 years. The successful cure percentage of retinal detachment critically relies on early-stage diagnosis. If addressed early, almost 90% of people with retinal detachment can recover from vision loss. Consequently, it is imperative to classify retinal detachment patients in an early phase. We developed a novel optimized lightweight multiclass retinal detachment grading model named LightMG-Net, which utilizes the image and feature-oriented handcrafted techniques for best feature analysis and the Grey Wolf Optimization technique for automatic hyperparameter tuning of a lightweight convolutional neural network for multiclass grading of retinal detachment from fundus images. The proposed LightMG-Net model is applied on four commonly utilized online databases, such as the Retinal Image Bank, Cataract Image Dataset, Kaggle, and Eye Disease Retinal Image for validation. The proposed LightMG-Net model achieved the best classification accuracy, sensitivity, specificity, and area under the curve at 95.42%, 95.10%, 98.90%, and 0.9947, respectively. Experimental outcomes demonstrate the superiority of the proposed approach over existing baseline methodologies.