Radiomic Features for Diabetic Retinopathy Early Detection Using Deep Neural Networks

Diabetic retinopathy is a common complication of diabetes and occurs when excess blood glucose damages the blood vessels of the retina. Early detection of this disease is crucial to prevent further damage to the retina and vision loss. However, in the early stages, the diagnostic process can be challenging as the findings in fundus images are less visible. Current diagnostic methods require some experience and ability from the doctor to achieve high efficiency. Therefore, there is a need for new approaches such as computer vision systems to assist ophthalmologists in detecting small areas that may contain abnormal signs such as hemorrhages and exudates. In this study, we address the early detection problem by developing a classification method for diabetic retinopathy using computer vision techniques. The proposed method extracted radiomic features from fundus images of patients diagnosed with mild diabetic retinopathy and a control group without eye disease. These features were then used to train a classifier based on deep neural networks. Our approach achieved an overall accuracy of 94.00%, a recall of 93.34%, and an area under the ROC curve of 97.07%. The findings of this study suggest that computer vision systems could be beneficially used in ophthalmic procedures for the early detection of diabetic retinopathy.