Prediction of ROP Zones Using Deep Learning Algorithms and Voting Classifier Technique

Abstract

Abstract Retinopathy of prematurity (ROP) can cause blindness in premature neonates. It is diagnosed when new blood vessels form abnormally in the retina. However, people at high risk of ROP might benefit significantly from early detection and treatment. Therefore, early diagnosis of ROP is vital in averting visual impairment. However, due to a lack of medical experience in detecting this condition, many people refuse treatment; this is especially troublesome given the rising cases of ROP. To deal with this problem, we trained three transfer learning models (VGG-19, ResNet-50, and EfficientNetB5) and a convolutional neural network (CNN) to identify the zones of ROP in preterm newborns. The dataset to train the model contains 1365 fundus images from the ROP screening. This dataset was gathered from the Private Clinic Al-Amal Eye center in Baghdad, Iraq. The models above are ensemble through voting classifier techniques to increase the performance. The proposed method had an overall accuracy of 88.82 percent when employing the voting classifier. On the other hand, EfficientNetB5 has outperformed other models in terms of accuracy with 87.27%.

Keywords

• Artificial intelligence
• Machine learning
• Deep learning
• Deep convolutional neural networks
• Voting classifier
• Retinopathy of prematurity