Pretrained Models and the Role of Feature Selection: An Artificial Intelligence-Based Approach in the Diagnosis of Diabetic Retinopathy

Abstract

Diabetic retinopathy is a significant complication occurring in the retina of the eye as a result of prolonged diabetes. When not detected early, this condition can lead to vision loss. Advanced image processing techniques and artificial intelligence algorithms have enhanced the possibilities of early diagnosis and treatment. This article discusses current advancements in artificial intelligence-based diabetic retinopathy detection and explores future possibilities in this field. In the experimental studies of the article, the Kaggle Aptos 2019 dataset was utilized. This dataset comprises 5 classes and a total of 3662 images. The class distribution is as follows: No DR (No Diabetic Retinopathy): 1805, Mild: 370, Moderate: 999, Severe: 193, Proliferative DR: 295. The study consists of four fundamental stages. These stages are (1) Feature extraction from VGG16 and VGG19 pretrained models, (2) Feature selection using NCA, Relieff, and Chi2, (3) Classification with Support Vector Machine classifier, (4) Iterative Majority Voting. Using the proposed method, a high accuracy of 99.18% is achieved. Furthermore, sensitivity of 100% for the No DR class, sensitivity of 100% for the Moderate class, sensitivity of 98.80% for the Severe class, and an F1-Score of 99.89% for the No DR class are obtained. This study demonstrates the effective utilization of machine learning methods in diabetic retinopathy diagnosis. The experimental results underscore the significant contributions of diabetic retinopathy patients' diagnosis and treatment processes.