Deep Learning for prediction of late recurrence of retinal detachment using preoperative and postoperative ultra-wide field imaging

Fiammetta Catania; Thibaut Chapron; Emanuele Crincoli; Alexandra Miere; Youssef Abdelmassih; William Beaumont; Ismael Chehaibou; Florence Metge; Sebastien Bruneau; Sophie Bonnin; Eric H Souied; Georges Caputo

doi:10.1111/aos.16693

Deep Learning for prediction of late recurrence of retinal detachment using preoperative and postoperative ultra-wide field imaging

Acta Ophthalmol. 2024 Apr 29. doi: 10.1111/aos.16693. Online ahead of print.

Authors

Affiliations

¹ Ophthalmology Department, Rothschild Foundation Hospital, Paris, France.
² Humanitas University, Department of Biomedical Sciences, Milan, Italy.
³ Université Paris Cité, CRESS, Obstetrical Perinatal and Paediatric Epidemiology Research Team, Paris, France.
⁴ Ophthalmology Unit, "Fondazione Policlinico Universitario A. Gemelli IRCCS", Catholic University "Sacro Cuore", Rome, Italy.
⁵ Department of Ophthalmology, Centre Hospitalier Intercommunal de Créteil, Créteil, France.

PMID: 38682863
DOI: 10.1111/aos.16693

Abstract

Purpose: To elaborate a deep learning (DL) model for automatic prediction of late recurrence (LR) of rhegmatogenous retinal detachment (RRD) using pseudocolor and fundus autofluorescence (AF) ultra-wide field (UWF) images obtained preoperatively and postoperatively.

Materials and methods: We retrospectively included patients >18 years who underwent either scleral buckling (SB) or pars plana vitrectomy (PPV) for primary or recurrent RRD with a post-operative follow-up >2 years. Records of RRD recurrence between 6 weeks and 2 years after surgery served as a ground truth for the training of the deep learning (DL) models. Four separate DL models were trained to predict LR within the 2 postoperative years (binary outputs) using, respectively, UWF preoperative and postoperative pseudocolor images and UWF preoperative and postoperative AF images.

Results: A total of 412 eyes were included in the study (332 eyes treated with PPV and 80 eyes with SB). The mean follow-up was 4.0 ± 2.1 years. The DL models based on preoperative and postoperative pseudocolor UWF imaging predicted recurrence with 85.6% (sensitivity 86.7%, specificity 85.4%) and 90.2% accuracy (sensitivity 87.0%, specificity 90.8%) in PPV-treated eyes, and 87.0% (sensitivity 86.7%, specificity 87.0%) and 91.1% (sensitivity 88.2%, specificity 91.9%) in SB-treated eyes, respectively. The DL models using preoperative and postoperative AF-UWF imaging predicted recurrence with 87.6% (sensitivity 84.0% and specificity 88.3%) and 91.0% (sensitivity 88.9%, specificity 91.5%) accuracy in PPV eyes, and 86.5% (sensitivity 87.5%; specificity 86.2%) and 90.6% (sensitivity 90.0%, specificity 90.7%) in SB eyes, respectively. Among the risk factors detected with visualisation methods, potential novel ones were extensive laser retinopexy and asymmetric staphyloma.

Conclusions: DL can accurately predict the LR of RRD based on UWF images (especially postoperative ones), which can help refine follow-up strategies. Saliency maps might provide further insight into the dynamics of RRD recurrence.

Keywords: artificial intelligence; deep learning; late recurrence; prediction; rhegmatogenous retinal detachment; ultra‐wide field imaging.