Prediction of Epiretinal Membrane from Retinal Fundus Images Using Deep Learning.

An epiretinal membrane (ERM) is an eye disease that can lead to visual distortion and, in some cases, to loss of vision. Screening retinal fundus images allows ophthalmologists to early detect and diagnose this disease; however, the manual interpretation of images is a time-consuming task. In spite of the existence of several computer vision tools for analysing retinal fundus images, they are mainly focused on the diagnosis of diabetic retinopathy and glaucoma. In this work, we have conducted a thorough study of several deep learning architectures, and a variety of techniques to train them, in order to build a model for automatically diagnosing ERM. As a result, we have built several models that can be ensembled to achieve a F1-score of 86.82%. The lessons learned in this work can serve as a basis for the construction of deep learning models for diagnosing other eye diseases.