Analysis of retinal images of premature infants using deep learning methods

Objective: The aim of this study is to develop a deep learning-based decision support system to assist clinical experts in the diagnosis of retinopathy of prematurity (ROP). In this context, retinal images of premature infants are analyzed using modern image processing techniques and deep learning algorithms, with the goal of automatically classifying the stages of ROP. The proposed system is intended to accelerate the diagnostic process, reduce reliance on expert interpretation, and enable early intervention to prevent the progression of the disease. Method: For this purpose, image optimization and standardization were achieved on retinal images with image processing techniques. Deep learning methods were used for the classification process. In particular, deep learning algorithms such as convolutional neural networks (CNN) were used to analyze the retinal images of premature babies and the classes of the disease (Normal-ROP0, ROP1, ROP2) were automatically detected. Results: There are a total of 13779 retinal images in the raw dataset obtained from Akdeniz University Faculty of Medicine. Then, the data in these images were filtered and multiplied. Finally, 25821 retinal images were obtained. An accuracy rate of %83,5 in AlexNet, %89,1 in Inception-V3, and %89,.9 in EfficientNet-B0 was determined for this dataset from CNN algorithms. Conclusion: Following image enhancement and deep learning-based classification, the highest performance was achieved by EfficientNet-B0, with an accuracy rate of %89,9. The AUC (Area Under the Curve) values for this model were determined to be 99% for the Normal-ROP0 class, 97% for ROP1, and 98% for ROP2. To further increase classification accuracy, it is recommended to enlarge the dataset and enhance the resolution and clarity of the retinal images.