Classification of Retinopathy of Prematurity using Neural Network

Retinopathy of prematurity typically diagnosed in prematurely-born infants. If it is not treated in time there is chances of retinal detachment and in advanced stages it may lead to permanent blindness. Identifying Retinopathy of prematurity with plus condition at early stage, may lead to start treatment as soon to get the best outcome. Here, we report on a computer-aided diagnosis (CAD) system to classify ROP affected vs. non-affected using retinal image captured using fundus camera. A dataset of 5000 images was collected including 900 with Plus or affected and 4100 non-affected images divided into training and testing sets. We tested logistic regression and neural network classifiers that achieved a mean AUC of 0.98 for the neural net as compared to logistic regression. The neural network was trained using a regularization parameter (λ) of 0.001 and step sizes of 0.001 and 0.005 for a total of 200 epochs. This shows that the neural network has a risk of miss diagnosis but they perform better than traditional algorithms showing its potentiality to assist in early accurate diagnosis of ROP as consequence providing clinicians with a powerful diagnostic and prognostic tool. However, these automated systems need further validation and optimization to verify that they can be used reliably in clinical practice.