Deep learning for detecting retinal detachment and discerning macular status using ultra-widefield fundus images

Abstract

Retinal detachment can lead to severe visual loss if not treated timely. The early diagnosis of retinal detachment can improve the rate of successful reattachment and the visual results, especially before macular involvement. Manual retinal detachment screening is time-consuming and labour-intensive, which is difficult for large-scale clinical applications. In this study, we developed a cascaded deep learning system based on the ultra-widefield fundus images for automated retinal detachment detection and macula-on/off retinal detachment discerning. The performance of this system is reliable and comparable to an experienced ophthalmologist. In addition, this system can automatically provide guidance to patients regarding appropriate preoperative posturing to reduce retinal detachment progression and the urgency of retinal detachment repair. The implementation of this system on a global scale may drastically reduce the extent of vision impairment resulting from retinal detachment by providing timely identification and referral.

Fig. 1

The workflow of developing deep learning (DL) systems for identifying retinal detachment (RD) and discerning macula-on/off RD based on ultra-widefield fundus images.

Fig. 2. Framework of the cascaded deep learning system and its corresponding clinical application.

The first model is used to identify retinal detachment (RD) and the second model is used to identify macula-on RD. UWF ultra-widefield fundus.

Fig. 3. Receiver operating characteristic (ROC) curves of the deep learning models derived from the test datasets, compared with general ophthalmologists’ performance using reference standard.

a Detection performance of retinal detachment. b Detection performance of macula-on retinal detachment. AUC area under the ROC curve; General ophthalmologist A, 5 years of working experience at a physical examination centre; General ophthalmologist b, 3 years of working experience at a physical examination centre.

Fig. 4. Ultra-widefield fundus images showing typical misclassified cases in retinal detachment (RD) detection.

a False-negative images: A1, RD surrounded by laser scars on the bottom left; A2, shallow RD at the bottom; A3, low-brightness image with RD on the top right. b False-positive images: B1, retinal breaks at the bottom; B2, subretinal membrane on the right side; B3, fundus albipunctatus.

Fig. 5. Ultra-widefield fundus images showing typical misclassified cases in macula-on retinal detachment (RD) detection.

a False-negative images: A1, macula-on RD with the distorted macula; A2, macula-on RD with an atrophic macula; A3, macula-on RD with an epiretinal membrane within the macular area. b False-positive images: B1, macula-off RD with a strong light reflex within the region of the macula; B2, shallow macular detachment; B3, exudative retinal detachment.

Fig. 6. Examples of ultra-widefield fundus images with an arrow/circle generated according to heatmaps.

a The arrow towards the area of retinal detachment in image A1 is established automatically based on the highlighted region in heatmap A2. b The circle located at the centre of retinal detachment in image B1 is created automatically on the basis of the highlighted region in heatmap A2. The arrow/circle is used to instruct patients in preoperative posturing to reduce the progression of retinal detachment between detection and treatment. The dotted diagonal line in the image is used to divide the retina into four quadrants (superior, inferior, left, and right).

Fig. 7. Example of an ultra-widefield fundus image with a mistaken circle generated according to a heatmap.

A circle which should have been a downward arrow in image A1 of inferior retinal detachment is established mistakenly based on the highlighted region in heatmap A2. The arrow/circle is used to instruct patients in preoperative posturing to reduce the progression of retinal detachment between detection and treatment. The dotted diagonal line in the image is used to divide the retina into four quadrants (superior, inferior, left, and right).