Diabetic retinopathy detection through convolutional neural networks with synaptic metaplasticity

Abstract

Background and objectives: Diabetic retinopathy is a type of diabetes that causes vascular changes that can lead to blindness. The ravages of this disease cannot be reversed, so early detection is essential. This work presents an automated method for early detection of this disease using fundus colored images. Methods: A bio-inspired approach is proposed on synaptic metaplasticity in convolutional neural networks. This biological phenomenon is known to directly interfere in both learning and memory by reinforcing less common occurrences during the learning process. Synaptic metaplasticity has been included in the backpropagation stage of a convolution operation for every convolutional layer. Results: The proposed method has been evaluated by using a public small diabetic retinopathy dataset from Kaggle with four award-winning convolutional neural network architectures. Results show that convolutional neural network architectures including synaptic metaplasticity improve both learning rate and accuracy. Furthermore, obtained results outperform other methods in current literature, even using smaller datasets for training. Best results have been obtained for the InceptionV3 architecture with synaptic metaplasticity with a 95.56% accuracy, 94.24% F1-score, 98.9% precision and 90% recall, using 3662 images for training. Conclusions: Convolutional neural networks with synaptic metaplasticity are suitable for early detection of diabetic retinopathy due to their fast convergence rate, training simplicity and high performance.