Improved Blood Vessels Segmentation of Infant Retinal Image.

Retinopathy of prematurity (ROP), is the most common cause of blindness in premature infants. ROP is measured by looking at the width, curvature, and length of the blood vessels map on a retina. So, the quality of the segmented blood vessel map affects how well the quantitative method works. Current vessel segmentation algorithms work well on images of the retina of adults, but they cannot tell the difference between structures of vessels that have not yet grown in images of the fundus of infant. Also, the lack of a dataset of infant fundus images has made it harder to develop data-driven techniques for separating blood vessels. This study shows how to use a Deep Convolutional Neural Network (DCNN)-based vessel segmentation system to determine if a infant has ROP. The proposed method uses a DCNN, Generative Adversarial Network (GAN) Pix2Pix, or U-Net to segment vessels. We trained the proposed system with datasets of fundus images that were available to the public, and we tested it with images of premature infants’ eyes from a nearby hospital. Experimental results show that the proposed method is more robust to noise and inter-class variation. It has a dice coefficient between 0.60 and 0.64 and an average accuracy of 96.69% for vessel segmentation. We have also examined its potential use in the treatment of ROP and Plus disease.