Deep Learning Models and Fusion Classification Technique for Accurate Diagnosis of Retinopathy of Prematurity in Preterm Newborn

Abstract

  Retinopathy of prematurity (ROP) is the most common cause of irreversible childhood blindness, and its diagnosis and treatment rely on subjective grading based on retinal vascular features. However, this method is laborious and error-prone, so automated approaches are desirable for greater precision and productivity. This study aims to develop a deep learning-based strategy to accurately diagnose the plus disease of ROP in preterm newborns using transfer learning models and a fusion classification technique. The Private Clinic Al-Amal Eye Center in Baghdad, Iraq, provided us with 2776 ROP screening fundus images between 2015 and 2020, and the images were used to train three deep convolutional neural network models (ResNet50, Densenet161, and EfficientNetB5). A fusion classifier approach was used to merge the three models for a thorough and precise diagnosis. The three models have relative accuracy rates of 69.78%, 80.57 %, and 81.29 % in their respective classifications. The overall accuracy, however, increased to 90.28 percent when the fusion classifier was employed. This shows that the proposed method helps identify ROP in premature infants. The study's findings imply the proposed method has the potential to significantly enhance the precision and speed with which ROP is diagnosed, which in turn could lead to earlier detection and treatment of the illness and a decreased likelihood of childhood blindness.