Multiclass Classification of Ocular Toxoplasmosis From Fundus Images With Residual Neural Networks

Ocular toxoplasmosis (OT) is usually diagnosed by a specialist through the analysis of fundus images of the eye. Deep learning has been used to perform binary classification of the disease; however, to the best of our knowledge this is the first work that performs a multi-class classification to differentiate between active or inactive OT lesions which gives more information to the clinician when making the diagnosis. Using a predictive model to differentiate healthy eyes from others with different type of lesions can be useful to save time, help diagnose atypical cases and assist specialists with less experience. In this work, we investigate the application of a deep learning model to perform multi-class image classification of eye fundus images. A pretrained residual neural network is fine tuned on a dataset of samples collected at the Hospital de Clínicas and Hospital General Pediátrico Acosta Nu medical centers from Asunción, Paraguay. The proposed model results are highly promising, scoring sensitivity and specificity rates equal to 91% and 98% respectively. In order to replicate the results and continue researching this area an open data set of images of the eye fundus labeled by ophthalmologists is also published as part of this work.