Machine Learning and Deep Learning Approaches for Exudate Segmentation and Classification in Diabetic Retinopathy

Diabetic retinopathy (DR) is the leading cause of blindness worldwide and refers to progressive degeneration of the retina. Early diagnosis and treatment of this condition are important to prevent permanent damage. Lipid deposits in the retina, which are found as exudates, are one of the major indicators of DR. The accuracy of detection, however, is affected by several factors, such as noise and image resolution. Deep learning (DL) models require significant computational power and vast quantities of data, whereas traditional machine learning methods are not capable of dealing with these challenges. This study proposes a fusion of machine learning and DL techniques for the automation of exudate segmentation and classification. The method applies preprocessing, which reduces noise and enhances input quality. A convolutional neural network (CNN) with an encoder-decoder structure based on U-Net performs pixel-wise exudate segmentation with precise localization. We applied a DL technique with pretrained DenseNet models to classify the exudates as mild, moderate, and severe. Using techniques such as rotation, zooming, and flipping, we enhance the model’s capability and help with data shortages. The standard datasets of digital retinal images for vessel extraction (DRIVE), structured analysis of the retina (STARE), diabetic retinopathy database (DIARETDB1), and IDRiD were used to evaluate the model. The results of the study show segmentation accuracy, measured by dice, of 0.91 (training), 0.89 (test), and 0.88 (validation). Furthermore, the classification shows an accuracy of 0.92 (training), 0.90 (test), and 0.88 (validation). Finally, both sensitivity and specificity were above 85%. The combination of preprocessing and transfer learning reduces the computational burden and makes it clinically usable in real-time.