Research in Medicine

Background and Aim:Diabetic retinopathy (DR) is a common microvascular complication of diabetes and a leading cause of blindness worldwide, resulting in visible structural changes in the retina. Early-stage detection significantly improves the chances of treatment and vision preservation. Recent advances have highlighted the role of artificial intelligence (AI), particularly deep learning, in the diagnosis of DR. This study aimed to evaluate the effectiveness and accuracy of AI-based methods in determining the severity of diabetic retinopathy.
Methods:A publicly available dataset containing 13,673 retinal images from 9,598 patients—originally compiled by Taveli et al. in 2019 under the supervision of seven physicians—was used to train convolutional neural networks (CNNs). The images, categorized into six classes based on disease severity and image quality, were often characterized by noise and low contrast, especially those captured using a fluorescent oscilloscope. To address this, we applied a contrast enhancement technique using constrained adaptive histogram equalization. To further improve model performance, the Archimedes metaheuristic optimization algorithm was employed to fine-tune CNN parameters. The algorithm was run with an initial population of 20 and 50 iterations.
Results:By combining image preprocessing with parameter optimization via the Archimedes algorithm, the proposed CNN model achieved an accuracy of 79% and a precision of 82% in classifying images into six severity categories. The optimized preprocessing and parameter tuning contributed significantly to enhancing the model’s performance.
Conclusion:This study demonstrates that accurate feature extraction from retinal images, coupled with intelligent parameter optimization using metaheuristic algorithms, can significantly improve the performance of AI-based DR diagnosis systems. The proposed model presents a promising alternative to existing CNN-based methods for classifying the severity of diabetic retinopathy.