Deep Learning for prediction of rhegmatogenous retinal detachment relapse using preoperative and postoperative wide field imaging

Abstract

Purpose : Rhegmatogenous retinal detachment (RRD) is complicated by relapse in 10-15% of cases. The aim of the study is to elaborate a deep learning (DL) model that would allow automatic prediction of 5-year relapse risk using color and autofluorescence (AF) ultra-wide field (UWF) fundus images taken prior to and immediately after surgery

Methods : We retrospectively searched for patients>18 years old treated with either scleral buckling (SB) or pars plana vitrectomy (PPV) for primary or relapsing RRD. Exclusion criteria were overlapping retinal diseases, syndromic RRD and follow up<1 year. Color and AF UWF fundus images taken preoperatively and postoperatively after tamponade resorption were collected and classified according to the occurrence of a relapse within 5 years. Images were cropped and used for training and internal validation (30% of the dataset) of 4 independent Deep Learning (DL) models.Testing performance of the DL models was evaluated in terms of sensitivity, specificity, accuracy and area under the ROC curves (AUROC)

Results : A total of 248 patients (170 PPV and 78 SB) were recruited. High myopia (>6 diopters) was present in 39 patients(16%). Age and sex were comparable between relapse and no relapse group. The DL model based on preoperative color UWF fundus images was trained and validated on 563 images and tested on 130 images; testing accuracy was 83.8% and 86.0% for PPV and SB respectively. The DL model based on postoperative color UWF fundus images was trained and validated on 578 images and tested on 138 images; testing accuracy was 89.0% for PPV and 91.1% for SB. The DL model based on preoperative AF reached a testing accuracy of 87.3% for PPV and 83.0% for SB. The DL model based on postoperative FA showed a testing accuracy of 92.2% for PPV and 87.5% for SB

Conclusions : Deep learning can accurately predict RRD relapse within 5 years from surgery. Postoperative UWF images may provide more accurate prediction compared to preoperative images. Color UWF imaging was a better predictor for relapse after SB surgery while UWF AF was a better predictor after PPV surgery

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.