Diagnosis of Diabetic Retinopathy with Transfer Learning and Metaheuristic Algorithms

Diabetic retinopathy is a common complication of diabetes, affecting a significant proportion of individuals with diabetes. The IDF Diabetes Atlas projects that the number of adults with diabetes will rise to 700 million by 2045. With the increasing prevalence of diabetes, it is expected that the number of individuals with diabetic retinopathy will also continue to rise. This highlights the importance of early detection, timely interventions, and effective management of diabetic retinopathy to prevent vision loss and other complications. Machine learning has the potential to enhance diabetic retinopathy (DR) diagnosis by providing efficient, objective, and consistent analysis of retinal images, improving early detection and management of the disease. This study proposes a hybrid model that applies deep learning and metaheuristic algorithms for diagnosing DR. A deep learning model based on CNNs, InceptionV3, is used for feature extraction from fundus images. The transfer learning method is applied in the extraction process. Two Metaheuristic algorithms: Particle Swarm optimization and Artificial Bee Colony, are applied for feature selection. Extracted and selected features are classified with eXtreme Gradient Boosting (XGBoost). This study shows that machine learning techniques with metaheuristic algorithms satisfactorily support ophthalmologists in diagnosing DR.