An Early Detection of Retinopathy of Prematurity on Fundus Image using Deep Learning

Abstract

Retinopathy of Prematurity (ROP) is one of the major causes of blindness in preterm babies, and early screening is important to prevent the loss of vision forever. Conventionally based diagnostic procedures depend on manual screening by ophthalmologists, which is time-intensive and demands expertise and professionalism. This study suggests an automated machine learning-based early ROP detection pipeline built on convolutional neural networks (CNNs) and transformers. The fundus image dataset was trained using the model with data augmentation methods, to adjust the imbalance in classes and increase generalization. Focal loss served to make the models pay attention to hard-to-detect cases of ROP at an early stage. The models were tested on accuracy, precision, recall, F1-score and Area Under the Curve (AUC). The findings indicated that the single models, i.e. the ResNet50 (85.84%), the ResNet101 (89.38%), the MobileNetV2 (85.84%), the EfficientNetV2 (91.15%), the ViT-B (88.12%), and the DeiT (89.09%), worked well, but the ensemble model between the ResNet101 and DeiT had the highest accuracy of 94.69%. The proposed system is a validated and automated system to detect ROP in healthcare, which may benefit healthcare professionals, particularly in lowresource environments to save on diagnostic time and increase accuracy, eventually leading to improved patient outcomes and alleviating the worldwide burden of ROPinduced blindness.