Predicting Diabetes from The Eye: A Retinal Imaging and Deep Learning Study

Abstract the primary objective of this cross-sectional study is to detect Diabetic Retinopathy (DR) in individuals who have undergone retinal imaging and eye examinations. To achieve this, the study employs customized retinal images and utilizes two machine learning techniques—Optimum-Path Forest (OPF) and Restricted Boltzmann Machine (RBM)—to classify images based on the presence or absence of DR. Feature extraction was carried out using both OPF and RBM models. In particular, the RBM model was trained to extract 500 and 1000 features from the retinal images after a thorough training phase. The evaluation process involved 15 separate experimental runs, each repeated 30 times to ensure reliable outcomes. The dataset included 73 diabetic patients, totaling 122 eyes examined. Participants had an average age of 59.7 years, with a nearly equal gender distribution (50.7% male and 49.3% female). Among the different model configurations tested, the RBM model with 1000 features (RBM-1000) emerged as the top performer, achieving an overall diagnostic accuracy of 89.47%. Furthermore, both the RBM-1000 and OPF-1000 models demonstrated perfect specificity, correctly identifying all images that did not show signs of DR. These results highlight the potential of machine learning techniques—especially the RBM model—in the automated detection of diabetic retinopathy. The high accuracy and specificity of the RBM approach suggest it could be a valuable tool for improving DR screening efficiency and supporting early diagnosis in clinical settings. Keywords: CNN, VGG-16, VGG-19, Feature Extraction, Localization, Segmentation, Region of Interest, Irido-diagnosis.