Disease Evaluation Of Retinopathy Of Prematurity(ROP) Based On Deep Neural Network

Retinopathy of prematurity(ROP)is a fundus vascular disease that mainly affects preterm infants less than 37 weeks of gestational age and low birth weight infants,and is one of the leading causes of blindness in infants.With the improvement of infant care,the prevalence of ROP has risen sharply in developing countries represented by China.Timely screening and intervention can greatly reduce the risk of blindness caused by ROP.However,the diagnostic criteria for ROP are complex,the testing equipment is expensive,and the clinical experience of professional ophthalmologists is highly dependent.Children in remote and underdeveloped areas cannot receive timely diagnosis and treatment.The development of computer technology provides technical support for the development of computer aided system and the introduction of expert experience to complete the diagnosis and treatment of ROP.However,due to the lack of fundus ROP images and the difficulty of labeling,most of the existing studies are still highly dependent on manual labor,and the diagnosis of ROP only focuses on determining the presence or absence of ROP,without further exploration.Convolutional neural networks are widely used in machine vision tasks such as image classification and segmentation because of their superior feature capture ability.The disease assessment method based on deep neural network can help doctors assess the severity of ROP more quickly and accurately,and then develop more accurate treatment plan.The lack of datasets has been an obstacle to the development of AI ophthalmic medicine.In this paper,over 70,000 fundus images of newborns were collected from Nanchang First Hospital and Ji’an Maternal and Child Health Hospital and annotated according to the international ROP staging standards to establish a ROP fundus image quality assessment dataset and a staging diagnosis dataset.In addition,due to the special nature of ROP fundus image collection,most images have poor contrast and unclear blood vessels.The CLAHE algorithm was applied to pre-process the images and highlight the lesion areas.Focusing on the early diagnosis of ROP,this paper aims to develop an end-to-end automatic diagnosis system to complete the diagnosis of ROP stage 1-3.The system consists of two parts.The first part is the ROP fundus image quality assessment network,which is responsible for screening out poor quality images at the source,so as not to affect the subsequent diagnosis results;The second part is the ROP staging diagnosis network,which gives the corresponding staging results for each input image.The ROP fundus image quality evaluation network selected lightweight convolutional neural network Mobile Net＿v2 as the backbone network,achieving94.54% accuracy,0.9478 F1 value and 0.9855 AUC value.The number of model parameters is 13.37 M,enabling rapid quality assessment.The ROP staging network is composed of three parts: backbone network,attention module and classifier.After comparison experiment,Dense Net-121 was selected as the backbone network for feature extraction.In order to improve the performance of staging network,a hybrid attention module(HAB)is proposed.This module can be integrated into any superior skeleton network to improve the performance of staging diagnosis for ROP.The network achieved F1 value of 0.8997,accuracy of 90.14% and Kappa value of 0.8577 on the test set,which has a good clinical application prospect.