Detection of Peripheral Retinal Breaks in Ultra-widefield Images Using Deep Learning

Retinal tears and holes are breaks in the retinal tissue that allow fluid to enter the subretinal space causing rhegmatogenous retinal detachment, a serious condition that if not treated in a timely manner produces visual impairment. In this paper, we propose a method to assist in the diagnosis of peripheral retinal breaks using Optomap Ultra-widefield (UWF) images and deep learning (DL) techniques. DL requires large amounts of labeled data, a difficult requirement to comply with. This article reports an experiment using a dataset with only 132 images, together with data augmentation and transfer learning techniques to overcome the problem of dataset size. We evaluate two pre-trained convolutional neural network models, AlexNet and SqueezeNet, with three different mini-batch sizes 8, 16 and 32. SqueezeNet mini-batch size 16 reached the best accuracy of 92.5 % in the test set. We demonstrated that our method is reliable for the early detection of retinal breaks and can be a useful addition to computer-aided diagnosis systems.