Deep Learning for Automatic Detection of Recurrent Retinal Detachment after Surgery Using Ultra‐Widefield Fundus Images: A Single‐Center Study

Abstract

It is important to detect recurrent retinal detachment (RD) among patients after retinal reattachment surgery. The application of deep learning (DL) methods to detect recurrent RD with ultra-widefield (UWF) fundus images is promising, but the feasibility and efficiency have not been studied. A DL system with ResNet-50 and Inception-ResNet-V2 is developed and internally validated to identify recurrent RD and retina reattachment after surgery. The performance is further validated and compared with human ophthalmologists in a prospective dataset assessed by area under curve (AUC), accuracy, sensitivity, and specificity. Five hundred fifty-four UWF fundus images from 173 RD patients (mean [standard deviation] age: 39.2 +/- 16.2 years; male: 115 [66.5%]) are used to develop the DL system. DL shows AUCs of 0.912 (95% confidence interval [CI]: 0.855-0.968) and 0.906 (95% CI: 0.818-0.995) for the two models. Eighty-nine UWF fundus images from 23 RD patients (mean [standard deviation] age: 31.4 +/- 12.3 years; male: 15 [65.2%]) are collected as prospective dataset. DL also shows the ability to detect recurrent RD with the AUCs of 0.929 and 0.930 for the two models, respectively. DL reaches a similar and even better diagnostic performance than junior ophthalmologists and performs much better than medical students.