Fine-tuning ResNet-50 for the classification of visual impairments from retinal fundus images

Abstract

The sense of sight plays a crucial role in human perception, as it serves as our primary sensory organ for perceiving light. However, a considerable number of individuals experience a wide range of vision impairments. These impairments encompass diverse conditions such as diabetic retinopathy, glaucoma, and cataracts. Each visual impairment exhibits unique characteristics and symptoms, highlighting the need for timely and accurate detection to facilitate appropriate treatment and prevent vision loss. This research aims to develop a deep learning-based system specifically designed to detect visual impairments. The proposed solution involves creating a model using the ResNet-50 algorithm as the foundational methodology, and fine-tuning multiple parameters to enhance the model's performance. The research utilizes a dataset consisting of retinal fundus images, which are categorized into four distinct classes: diabetic retinopathy, glaucoma, cataracts, and normal. The findings demonstrate the effectiveness of the model, achieving an impressive accuracy score of 92%. This signifies a significant improvement of 6% over the accuracy achieved in the previous study, which stood at 86%. The implementation of this system is expected to make a significant contribution to the rapid and accurate detection of various eye disorders in the future, enabling timely intervention and prevention of visual impairment.