A cascaded deep learning system for detecting retinal detachment and discerning macular status using ultra-widefield fundus images

Abstract

Purpose : Retinal detachment (RD) can lead to severe visual loss if not treated timely. The early diagnosis of RD can improve the rate of successful reattachment and the visual results, especially before the macular involvement. Manual RD screening is time-consuming and labor-intensive, which is difficult for large-scale clinical applications. This study aimed to develop and evaluate a cascaded deep learning (DL) system for automated RD detection and macula-on/off RD discerning based on ultra-widefield fundus (UWF) images.

Methods : A cascaded DL system including two different models was developed using 11087 UWF images from 7966 individuals. The first model is used to detect RD, and the second model is used to discern macula-on RD from macula-off RD. This study recruited 3 experienced retinal specialists to classify the images separately. The reference standard was determined when the agreement was achieved among all 3 retinal specialists or arbitrated by another senior retinal specialist if any disagreements existed. Two independent test sets were used to assess the performance of the DL models. Heatmaps were generated to highlight the regions that the DL model used to identify RD. Area under the receiver operating characteristics curve (AUC), sensitivity and specificity of the DL models were evaluated and compared with the performance of ophthalmologists.

Results : The first DL model used to identify RD achieved an AUC of 0.989 (95% CI: 0.978-0.996), with a sensitivity and specificity of 96.1% and 99.6%, respectively. The second DL model used to discern macula-on RD from macula-off RD achieved an AUC of 0.975 (95% CI: 0.957-0.988), with a sensitivity and specificity of 93.8% and 90.9%, respectively. The performance of this system is comparable to the ophthalmologist with five years of clinical experience. All 292 true positive RD images in the test set displayed heatmap visualization in the RD regions. The instruction regarding appropriate preoperative posturing to reduce RD progression was automatically provided by the system according to the RD location.

Conclusions : A fully data-driven cascaded DL system with two models can be used to detect RD and discern macula-on/off RD with high reliability. The implementation of this system on a global scale may drastically reduce the extent of vision impairment resulting from RD by providing timely identification and referral.

This is a 2020 ARVO Annual Meeting abstract.