A Systematic review on using deep learning in Retinopathy of Prematurity

Background and aims: Retinopathy of prematurity (ROP) is an eye disease that can happenin premature babies. It causes abnormal blood vessels to grow in the retina and can lead to visionproblems and blindness. Retinopathy of prematurity can be a treatable disease with an appropriateand timely diagnosis. In the last two decades, many different approaches have been applied inRetinopathy of prematurity detection. A lot of AI algorithms have been developed for the detectionand diagnosis of diseases such as cancer, heart disease, and Parkinson’s disease. In ophthalmology,it has demonstrated clinically acceptable diagnostic performance in detecting diabeticretinopathy (DR), glaucoma, and age-related macular degeneration using fundus photos and OCT.Compared to traditional computation algorithms, machine learning, and deep learning algorithmsare more effective in diagnosis and disease detection. This study aimed to review the diagnosticaccuracy of deep learning algorithms to identify retinopathy of prematurity.Method: A systematic search was conducted in Medline (PubMed), Scopus, and Web of Science(WOS) using a combination of the key terms retinopathy of prematurity and deep learning andtheir alias from January ۲۰۱۵ to February ۲۰۲۳. The title, abstract and full text of the extracted articleswere then viewed using the PRISMA checklist. Studies dealing with the application of deeplearning algorithms in connection with retinopathy of prematurity were included. The informationregarding developed models was extracted from reviewed articles.Results: Of ۱۶۳۱ searched articles between ۲۰۱۵ and ۲۰۲۳, ۳۵ studies met the inclusion criteria.All studies focused largely on using deep learning to detect plus disease in retinopathy of prematurityscreening or to determine ROP severity level or both. Moreover, there were ۸ availablepremature infants’ retinal image datasets for ROP using for detection, classification, and segmentation.Studies showed the area under the curve (AUC)’s ranged between ۰.۸۳۴ and ۰.۹۹۴for diagnosing retinopathy of prematurity on retinal images. The performance of deep learningalgorithms was excellent and promising. This review showed that deep neural networks (DNNs)have become the most favored and approved method for ROP detection. Among these DNNmethods, Convolutional Neural Network (CNN) models were the most frequently applied in theclassification of medical image data.Conclusion: DL approaches provide enhanced automation which accelerates evaluation and goesbeyond traditional statistical analysis. Considering the high performance that DL models haveshown in the ROP context, these techniques can be a good alternative to the traditional methodsfor diagnosing this disease.

کلیدواژه ها:

artificial intelligence ، deep learning ، retinopathy of prematurity ، convolutional neural network