A Deep Learning Approach For Prediction For Retinopathy Of Prematurity And Related Complications

BackgroundRetinopathy of prematurity(ROP)is a retinal vasoproliferative disease that affects preterm infants and low-birth-weight infants.It is one of the leading causes of childhood blindness worldwide.Regular screening,early diagnosis,and timely treatment can prevent visual impairment.Prediction models for ROP that could identify infants at high risk of ROP might safely reduce the number of unnecessary screening examinations.ROP is often associated with other premature birth complications,thus fundus images of infants may be a window to understand these diseases.Prediction models can be constructed by using deep learning to mine the image features of ROP and other related premature birth complication from the fundus images of infants,to provide an individualized ROP screening strategy and clinical management scheme for infants.Objective1.This study aimed to develop and validate a deep learning prediction model for ROP and type 1 ROP based on the fundus images at the first screening examination.2.This study aimed to develop and validate a deep learning prediction model for ROP related complications,including bronchopulmonary dysplasia,patent ductus arteriosus,necrotizing enterocolitis,and intraventricular hemorrhage,based on the fundus images at the first screening examination.MethodsThis study retrospectively collected 1020 preterm infants who underwent ROP screening from June 2017 to November 2020 in Zhujiang Hospital of Southern Medical University for model training and internal validation.In addition,85 preterm infants who were screened for ROP from September 2019 to August 2020 in Nanning Second People’s Hospital were collected for external validation.The EfficientNet-B4 convolutional neural network architecture was used for transfer learning to construct disease prediction models.The accuracy,sensitivity,specificity,positive predictive value,negative predictive value,and F1 score,were calculated to evaluate the performance of the predictive models.Results1.For predicting ROP,the prediction model demonstrated an average sensitivity of 0.89 and an average specificity of 0.86 in the five-fold cross-validation.In the external validation,the sensitivity was 0.88,and the specificity was 0.86.For predicting type 1 ROP,the prediction model demonstrated an average sensitivity of 0.85 and an average specificity of 0.84 in the five-fold cross-validation.In the external validation,the sensitivity was 0.75,and the specificity was 0.79.2.For predicting bronchopulmonary dysplasia,the prediction model demonstrated an average sensitivity of 0.74 and an average specificity of 0.83 in the five-fold cross-validation.In the external validation,the sensitivity was 0.81,and the specificity was 0.78.For predicting patent ductus arteriosus,the prediction model demonstrated an average sensitivity of 0.80 and an average specificity of 0.84 in the five-fold cross-validation.In the external validation,the sensitivity was 0.77,and the specificity was 0.78.For predicting necrotizing enterocolitis,the prediction model demonstrated an average sensitivity of 0.76 and an average specificity of 0.80 in the five-fold cross-validation.In the external validation,the sensitivity was 0.67,and the specificity was 0.77.For predicting intraventricular hemorrhage,the prediction model demonstrated an average sensitivity of 0.81 and an average specificity of 0.83 in the five-fold cross-validation.In the external validation,the sensitivity was 0.83,and the specificity was 0.81.Conclusion1.The deep learning model constructed based on convolutional neural network has encouraging predictive value for ROP and type 1 ROP.2.The deep learning model based on convolutional neural network showed encouraging predictive performance for bronchopulmonary dysplasia,patent ductus arteriosus,necrotizing enterocolitis,and intraventricular hemorrhage.