DIABETIC RETINOPATHY RECOGNITION USING IMAGE PROCESSING AND MACHINE LEARNING MODELS ON FUNDUS IMAGES

Abstract

This research develops a machine learning classification framework that integrates multiple image processing techniques and canonical image alignment to enhance diabetic retinopathy recognition from fundus photographs. Color fundus images from the Ocular Disease Intelligent Recognition (ODIR) dataset, collected by Shanggong Medical Technology Co., Ltd. from different hospitals/medical centers in China, were preprocessed using ridge detection, Sobel edge detection, Canny edge detection, Robert edge detection, and color contrast enhancement. Canonical fundus images were constructed to standardize structural representation across samples, and an additional dataset was created by resizing and realigning the processed images, using image registration techniques, to match their corresponding canonical images. Six classification methods, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Naïve Bayes, Decision Trees, Gradient Boosted Trees (GBT), and Wasserstein Distance Classification (WDC), were evaluated using training-testing splits of 70:30, 75:25, and 80:20. For left-eye fundus photographs, the best performance was obtained using GBT with a 75:25 training-testing split on canonical-aligned images. The model achieved an accuracy of 0.7715, a precision of 0.7637, a recall of 0.7715, and an F1-score of 0.7645. For right-eye photographs, the highest performance was also achieved by GBT, using a 70:30 training–testing split on canonical-aligned images, yielding an accuracy of 0.7196, a precision of 0.7125, a recall of 0.7196, and an F1-score of 0.7039. These findings indicate that aligning processed images to canonical fundus templates enhances structural consistency and improves the classifier’s ability to detect diabetic-related abnormalities. Overall, the results demonstrate that combining multiple image processing techniques with canonical image alignment substantially improves diabetic retinopathy classification performance. The evidence supports the potential of machine-learning-based frameworks as effective screening tools for early detection of diabetic retinopathy.