Automated Detection and Classification of Diabetic Retinopathy Stages Using SVM, Random Forest, and k-NN on Retinal Fundus Images

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness globally, which makes early detection not only crucial but also the most effective way to manage the disease. In this research, a machine learning (ML) model for automated DR identification and staging primarily using retinal fundus images was proposed. Classical ML methods, namely Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbors (k-NN), were utilized on the basis of image features such as texture, color, and luminance. The images were resized and normalized, and data was augmented to help the models learn better and generalize well to new data. Results from the experiments indicated that SVM and RF, besides k-NN, showed better accuracy, precision, and sensitivity, suggesting their effectiveness in addressing complicated classification problems. Accuracy, F1-score, and ROC-AUC were the additional metrics used to assess the models performance, which demonstrated consistent classification of different DR stages. The Brier Score (BS) and Expected Calibration Error (ECE) used for calibration checking also showed well calibrated probability forecasts. Although the study provided initial evidence supporting the efficacy of machine learning in DR screening, further research is needed to address the identified issues such as class imbalance and reliance on handcrafted features. In a nutshell, the paper demonstrates machine learning approaches as fast and cost effective automated DR screening instruments, presents a strong argument for the usage of such techniques, and brings to light the complexity of real world clinical implementation.