A Robust Deep Learning Detection Approach for Retinopathy of Prematurity

Retinal retinopathy of prematurity (ROP), an abnormal blood vessel formation, can occur in a baby who was born early or with a low birth weight. It is one of the primary causes of newborn blindness globally. Early detection of ROP is critical for slowing and stopping the progression of ROP-related vision impairment which leads to blindness. ROP is a relatively unknown condition, even among medical professionals. Due to this, the dataset for ROP is infrequently accessible and typically extremely unbalanced in terms of the ratio of negative to positive images and the ratio of each stage of it. This paper addresses the rarity of datasets and the difficulty of detecting ROP in retinal fundus images. Using our own collected dataset to handle the data problem, we then use state-of-the-art deep learning models with the use of transfer learning and some techniques to build a robust model with an accuracy of 96.64% that can help doctors in the diagnosis process which will lead to a great effect in the healthcare system regarding this problem.