Classification of Diabetic Retinopathy using Statistical Region Merging and Convolutional Neural Network

Abstract

Diabetes Retinopathy is one of the complications of Diabetes Mellitus, which occurs in the retina of the eye. Patients will experience vision problems and if it handled late, patients will experience blindness. Manual examination by an ophthalmologist will take time and the results of the examination also depend on the doctor's expertise in diagnosing. Therefore, digital image processing system which able diagnose quickly, accurately, and objectively, is needed. Based on these problems, in this research the author designed a system that can process digital fundus images and classify them into 4 classes namely normal, mild NPDR, moderate NPDR, and severe NPDR by using data taken from MESSIDOR dataset. The data is processed using Statistical Region Merging (SRM) segmentation method and is classified using the Convolutional Neural Network (CNN) method. The number of data samples used in this research is 80 images, which consist of 20 image samples for each class. The best accuracy achieved was 81.25% using a ratio of 3:2 training data and test data, the value of segmentation complexity parameters $\mathrm{Q}={256}$, the number of $\mathrm{epochs}={100}$ and the learning rate $\mathrm{e}={0.0001}$, with 14.518 seconds of computation time.