Abstract
Purpose :
To develop a deep learning model that can effectively detect referable and vision-threatening diabetic retinopathy (RDR and VTDR) on images obtained from ultra-wide field scanning laser ophthalmoscope (UWF-SLO).
Methods :
UWF-SLO (Daytona, Optos, Dunfermline, UK) images were retrospectively collected from subjects with diabetes in CUHK Eye Centre, Hong Kong for developing the deep learning algorithm. All images were labeled as yes/no RDR and yes/no VTDR by retinal specialists during fundus examination, according to the International Clinical Diabetic Retinopathy Disease Severity Scale. RDR was defined as moderate or above non-proliferative DR and/or presence of diabetic macular edema (DME), and VTDR was defined as severe non-proliferative DR, proliferative DR, and/or presence of DME. Convolutional Neural Network (ResNet50) classifiers were first trained for filtering ungradable images, and then classifying yes/no RDR and yes/no VTDR. One external dataset (External-1: 1321 images from 232 eyes of 232 subjects with diabetes) was used to test the classification of yes/no gradable image. Two external datasets (External-2: 3959 images from 648 eyes of 378 subjects with diabetes, External-3: 218 images from 216 eyes of 118 subjects with diabetes) were used to test the classification of yes/no RDR and yes/no VTDR.
Results :
In development stage, 4095 images from 1043 eyes of 524 diabetic subjects were included for training (60%), validation (20%) and primary testing (20%). In the primary testing, the model achieved AUROC of 0.923 and accuracy of 86.4% for detecting ungradable images. For detecting RDR and VTDR, the model achieved AUROC of 0.971 and 0.940, and accuracy of 92.8% and 88.8%, respectively. In the external validation, using the External-1 dataset, the model achieved AUROC of 0.911, and accuracy of 92.2% for detecting ungradable images. Using the External-2 dataset, the model achieved AUROC of 0.848 and 0.872, and accuracy of 74.0% and 78.1% for detecting RDR and VTDR, respectively. Using the External-3 dataset, the model achieved AUROC of 0.878 and 0.916, and accuracy of 82.3% and 69.4% for detecting RDR and VTDR, respectively.
Conclusions :
The proposed deep learning model could detect RDR and VTDR from UWF-SLO images with excellent performance.
This is a 2020 ARVO Annual Meeting abstract.