Early Detection of Premature Retinopathy using Deep Learning Techniques

Retinopathy of prematurity (ROP) can be cured with laser therapy or surgery, early identification is crucial. In most situations, treatment can avert blindness. On the other hand, ROP can result in irreversible blindness if it is not identified and treated quickly. Using InceptionV3, ResNet50, VGG16, VGG19, DenseNet201, MobileNet, AlexNet, and Ensemble Model, the proposed study seeks to classify various stages of Retinopathy of prematurity (ROP). Initially, 91 retinal fundus images were gathered, comprising no ROP, stage one, stage two, and stage three images. Subsequently, four discrete datasets were generated through the use of four image processing techniques: histogram equalization, canny filter, sobel filter, and enhanced multispectral colour composite. The dataset's size was raised from 91 to 1000 images using image augmentation techniques, with 250 images each class. After comparing the outcomes, it was discovered that, out of all the combinations, employing MobileNet in Contrast with Histogram equalisation and enhanced multispectral colour composite datasets produced the greatest test accuracy, at 99.85% and 95.55%, respectively.