Predicting Visual Outcomes in Congenital Posterior Lens Opacities: A Deep Learning Model Using Ultra-Widefield Fundus Images

Purpose: To develop and validate a deep learning (DL) model for the automatic segmentation of lens opacity projected shadow (LOPS) on ultra-widefield (UWF) fundus images and to predict visual outcomes in children with congenital posterior lens opacities (CPLOs). Design: A prospective single-center study. Participants: Children with CPLO who underwent surgery at the Zhongshan Ophthalmic Center in China from January 2023 to November 2023. Methods: Demographics, best-corrected visual acuity (BCVA), biomicroscopic assessments, and UWF fundus images were collected. A DL model based on U-Net architecture was implemented for the segmentation of total and dense LOPS and optic disc. The ratio of LOPS to optic disc was defined as shadow-to-optic disc ratio (SODR). The DL model was evaluated by confusion matrix and Dice coefficients. An optimal visual outcome was defined as a BCVA better than 0.30 logarithm of the minimal angle of resolution. Factors correlated with optimal visual outcomes were identified by regression analyses, and predictive performance was evaluated by receiver operating characteristic curve. Main Outcome Measures: Pixel area of total and dense LOPS, total and dense SODR, and postoperative BCVA at a minimum 12-month follow-up. Conclusions: The LOPS on UWF fundus images can be accurately segmented by DL model. A dense-SODR ≤14.0 is associated with optimal visual outcomes, offering a novel, valuable predictor for visual outcomes in CPLO. Financial Disclosures: The authors have no proprietary or commercial interest in any materials discussed in this article.