Abstract
Premature adolescents with Retinopathy of Prematurity (ROP), a
fibrovascular proliferative condition, have difficulties with the maturing peripheral
retinal vasculature. Early identification of ROP is achievable in stages 1 and 2,
distinguished by a demarcation line and ridge that divides the peripheral retina from the
vascularized retina. Because newborn retinal images have poor contrast, it is difficult to
distinguish demarcation lines or ridges. This study used segmentation and
convolutional neural networks to detect ridges, which are crucial landmarks in the
diagnosis of ROP. Our contribution is implementing Mask R-CNN for identifying
boundary line/ridge recognition, which enables doctors to identify ROP stage 2 more
accurately. To combat poor image quality, the suggested approach uses a pre-processing stage of image augmentation. In this study, the utility of the Convolutional
Neural Network was examined to localize ridges in labeled neonatal photos. The
KIDROP study and a dataset comprising 220 photos of 45 infants were used. Using the
segmentation of the ridge region as the ground truth, 175 retinal images were used to
train the system. The system's detection accuracy was 0.94, with 45 images under test,
proving that data augmentation detection in conjunction with image normalizing pre-processing allows accurate identification of the ROP in its early stages.