Optimizing Early Detection of Diabetes through Retinal Imaging: A Comparative Analysis of Deep Learning and Machine Learning Algorithms

Abstract

Early detection of diabetes is essential to prevent complications like diabetic retinopathy. This study evaluates the effectiveness of five algorithms—Convolutional Neural Networks (CNN), Random Forest, Support Vector Machines (SVM), Gradient Boosting, and K-Nearest Neighbors (KNN)—in detecting diabetes through retinal imaging. Using a dataset of 15,000 retinal images, models were assessed for accuracy, precision, recall, and F1-score, with image preprocessing and data augmentation such as Rotation (θ\thetaθ), Flipping (Horizontal and Vertical), Zooming (zzz), Brightness Adjustments, CLAHE, sharpening filters and Gaussian Blur are applied to enhance performance. CNN beat the other models, reaching a 97.2% accuracy, indicating its supremacy in predicting performance. The comparison research also showed distinct strengths and limits of each method, demonstrating their usefulness across diverse diagnostic circumstances. These results underline the transformational potential of machine learning in medical diagnostics, notably for retinal imaging-based diabetes identification.