Deep Learning for Diabetic Retinopathy Prediction

Abstract

Diabetic retinopathy is a complication of diabetes mellitus. Its early diagnosis can prevent its progression and avoid the development of other major complications such as blindness. Deep learning and transfer learning appear in this context as powerful tools to aid in diagnosing this condition. The present work proposes to experiment with different models of pre-trained convolutional neural networks to determine which one fits best the problem of predicting diabetic retinopathy. The Diabetic Retinopathy Detection dataset supported by the EyePACS competition is used for evaluation. Seven pre-trained CNN models implemented in the Keras library developed in Python and, in this case, executed in the Kaggle platform, are used. Results show that no architecture performs better in all evaluation metrics. From a balanced behaviour perspective, the MobileNetV2 model stands out, with execution times almost half that of the slowest CNNs and without falling into overfitting with 20 learning epochs. InceptionResNetV2 stands out from the perspective of best performance, with a Kappa coefficient of 0.7588.