Identification of Diabetic Retinopathy from Retinography Images Using a Convolutional Neural Network

Diabetic retinopathy (DR) is a prevalent eye disease in people with diabetes worldwide and can cause vision loss or blindness. Conventional diagnostic imaging requires time, effort and specific skills of ophthalmologists. This study proposes the use of a convolutional neural network (CNN) based on the ResNet152V2 architecture to automatically analyze color images of the retina of the eye and identify DR. The Knowledge Discovery in Databases (KDD) methodology was applied for data management and analysis. Datasets of RGB images were acquired, both private from the Ecuadorian Diabetes Association (EDA) and public (EyePACS) available on the Internet. Training and validation of the model were performed with Python, the TensorFlow framework and the Keras library. The results showed that the model has an accuracy in DR identification of 80% comparable to that of ophthalmologists (specialists), showing a statistically significant association according to the chi-square test and a very high Spearman correlation (rho = 0.857). This resulted in a high concordance between both evaluations (model vs. specialists). In addition, the CNN model significantly reduced the manual DR diagnosis time from 5-10 min to 15-30 s. The implementation of this tool could potentially improve the diagnosis of DR and the prescription of appropriate clinical treatments for affected patients.