Data augmentation for deep learning of non-mydriatic screening retinal fundus images

Abstract

Fundus image is an effective and low-cost tool to screen for common retinal diseases. At the same time, Deep Learning (DL) algorithms have been shown capable of achieving similar or even better performance accuracies than physicians in certain image classification tasks. One of the key aspects to improve the performance of DL models is to use data augmentation techniques. Data augmentation reduces the impact of overfitting and improves the generalization capacity of the models. However, the most appropriate data augmentation methodology is highly dependant on the nature of the problem. In this work, we propose a data augmentation and image enhancement algorithm for the task of classifying non-mydriatic fundus images of pigmented abnormalities in the macula. For training, fine tuning and data augmentation, we used the Barcelona Supercomputing Centre cluster CTE IBM Power8+ and Marenostrum IV. The parallelization and optimization of the algorithms were performed using Numba, and Python-Multiprocessing, made compatible with the underlying DL framework used for training the model. We propose and trained a specific DL model from scratch. Our main results are an increase in the number of input images up to a factor of, and report the information of quality images for. As a result, our data augmentation approach results in an increase of up to 9% in classification accuracy.