Detection Of RRD And Its Related Peripheral Retinal Lesions By Using Deep Learning

Purpose: To develop a deep learning system(DLS)that can automatically detect and categorise rhegmatogenous retinal detachment(RRD)and its related peripheral retinal lesions with ultra-widefield(UWF)fundus images.Methods: This study collected UWF fundus images of patients from Joint Shantou International Eye Center from November 2015 to January 2021.We set up four categories of lesions: cystic retinal tuft(CRT),lattice degeneration,retinal breaks and RRD.After completion of quality assessment and manual labeling,the datasets were randomly divided into three portions: a training dataset(75%),a validation dataset(10%)and a test dataset(15%),according to the number of patients.We explored deep learning algorithm with the ensemble of three CNN: Inception V3,Inception Res Net V2 and Xception,as well as multi-label classification,to establish a model for RRD and its related peripheral retinal lesions detection.Accuracy,sensitivity,specificity,F1 score and AUC were applied to evaluate the performance and clinical value of the deep learning model.In addition,visualization technique was adopted to explain the model.Results: 5,958 UWF images from 3,740 participants and 4,944 eyes were analysed.The numbers of CRT,lattice,retinal breaks and RRD were respectively 65,827,1,011 and 953.In the test dataset,for CRT diagnosis,the deep learning model showed AUC 0.9781(95%CI,0.9538-1.0000)and sensitivity 0.867;for lattice,AUC0.9550(95%CI,0.9358-0.9743)and sensitivity 0.881;for retinal breaks,AUC0.9205(95%CI,0.8989-0.9421)and sensitivity 0.836;for RRD,AUC 0.9882(95% CI,0.9758-1.0000)and sensitivity 0.918.Moreover,the referral performance of this model indicated F1 score 0.905 and accuracy 0.938.Conclusion: Based on UWF fundus images,the deep learning algorithm model for detection of RRD and its related peripheral retinal lesions built in this research showed good AUC and accuracy.