Optimizing Retinal Fundus Images for the Automated Diabetic Retinopathy Diagnosis

Diabetic retinopathy (DR) is a common eye condition arising from the complications associated with diabetes. The diagnosis of DR is traditionally carried out by experienced clinicians who screen fundus images to identify numerous small features. This process is challenging, time-consuming, and susceptible to errors. Various studies have successfully utilized different machine learning techniques for the automated diagnosis of DR. The performance of DR diagnosis can be influenced by the preprocessing of the images. In this study, the impact of unifying image resolution and field of view on the performance of machine learning-based automated DR diagnosis was evaluated. This approach ensured that the sizes of the various internal parts of the retina were effectively standardized across different images. Over 32,000 images from two different datasets obtained from kaggle.com were utilized. Two distinct classification experiments were conducted on the two datasets using an Inception convolutional neural network, both with and without the proposed method. The proposed method resulted in accuracies of 78.8% and 96.8% for dataset 1 and dataset 2, respectively, compared to accuracies of 78.1% and 95.9% without this method.