Computer Aided Diagnostics in Prediction of Plus and Pre Plus Diseases of ROP in Neonatal Fundus Images

Retinopathy of Prematurity (ROP) is the premature retinal disorder causing abnormal vasculatures in neonates. It is a condition in premature infants that has the potential to cause blindness. Timely diagnosis and treatment of the disease in premature infants is the only key to avoid permanent blindness in them. The application of deep learning in diagnosis of ROP aids in reducing the stressful manual diagnosis of ROP by the ophthalmologists. The proposed study is aimed to automatically diagnose Plus, Pre Plus diseases in the retinal images collected real time, retrospectively. Six pre-trained networks such as ResNet50, EfficientNet, ShuffleNet V2, DarkNet19, Alex Net and VGG16, are utilised in the classification of provided neonatal images into three classes namely, Plus, Pre Plus and normal using transfer learning approach. It is inferred that ResNet50 network outperforms all other used pre-trained networks with an classification accuracy of 96.7 %, sensitivity of 96.83 %, specificity of 96.67% and AUC of 0.964. The proposed study can be utilized in first line screening of thousands of premature infants in the diagnosis of ROP.