Abstract:
Diabetic Retinopathy (DR) is a serious complication of diabetes that can lead to blind
ness if not detected early. Diabetes, along with diabetic retinopathy, is associated with
causing harm to the eyes, characterized by the leakage of blood vessels damaging the
retina, leading to microaneurysms (MA), hard exudates (HE), and soft exudates (cot
ton wool spots). As a chronic condition, DR poses a growing challenge worldwide,
often resulting in visual impairment or blindness if left undiagnosed. Early detection
is crucial, yet challenging, as symptoms manifest only after signi cant retinal changes
occur. This study investigates the e ectiveness of a multimodal approach using ma
chine learning classi ers for the early detection of DR, employing both binary and
multiclass classi cation methods. We utilized K-Nearest Neighbors (KNN), Support
Vector Machine (SVM), and ResNet-50 algorithms to classify retinal images. In binary
classi cation, designed to distinguish between DR and non-DR images, KNN achieved
an accuracy of 88%, SVM reached 92%, and ResNet-50 outperformed both with an
accuracy of 98%. These results demonstrate the potential of machine learning, par
ticularly deep learning with ResNet-50, in accurately identifying the presence of DR.
For multiclass classi cation, which aims to classify images into di erent stages of DR
(no DR, mild, moderate, severe, and proliferative DR), KNN achieved an accuracy of
70%, SVM achieved 71%, and ResNet-50 achieved the highest accuracy of 82%. The
study's ndings highlight the superior performance of ResNet-50 in both binary and
multiclass classi cation tasks, suggesting its suitability for automated DR screening
systems. The implementation of such advanced machine learning models can aid in
early diagnosis, thus facilitating timely treatment and potentially preventing vision loss
in diabetic patients.
