Utilizing Machine Learning Algorithms for Diagnosing Diabetic Retinopathy

Diabetic Retinopathy (DR) is a prevalent eye disease that affects up to 80% of the diabetic population worldwide and it is the second leading cause of blindness in the world. DR has become an important cause of visual impairment among people in ages of 25-74 years old. A precise detection and intervention of DR are crucial, as they can help mitigate the medical and economic impact of the disease. When timely treatment is provided to the DR patients, approximately 90% of patients can be saved from visual loss. However, the current diagnosis process is time-consuming and requires specialized expertise, which might be costly. This paper introduces a new automated model for detecting and classifying DR from color fundus images. The images are preprocessed and resampled to ensure a balanced dataset, which is essential for the trained model to treat all categories equally. Afterward, the system combines two feature extraction techniques: Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) to extract the required features from the preprocessed image. These features are optimized using mutual information techniques to create an optimal feature set. It subsequently undergoes classification using different classification models: SVM, Random Forest, and Gradient Boosting. A series of experiments take place on APTOS and MESSIDOR datasets. The highest accuracy of 99% on both datasets demonstrates the effectiveness and generalizability of the proposed approach.