Abstract
Purpose :
To evaluate the performance of a deep learning algorithm for the automated grading of vitritis on ultra-wide field imaging.
Methods :
This was a cross-sectional non-interventional study. Ultra-wide field (UWF) fundus retinophotographs (Optos®, Dunfermline, Scotland, United Kingdom) of consecutive patients followed for intermediate, posterior or pan-uveitis were used. Vitreous haze was defined on each UWF picture by two blinded experts according to the 6 steps of the Nussenblatt scale. Included images were automatically classified using the Inception V3 convolutional neural network from Google (Mountain View, California, USA). Images were fed to the model as follows: training (70%), validation (15%), and testing set (15%). The performance of the system was assessed on the unused testing set.
Results :
A total of 1181 images from 443 patients were included. There were 55% of females and the mean age was 51.9 years (range: 9-92, SD: 18). Sixty percent (N=708) of patients were Caucasian, 21% (N=248) of North African, 13% (N=154) of African, and 6% (N=71) of Asian heritage respectively. The main etiologies of uveitis were idiopathic 45% (N=532), sarcoidosis 13% (N=154), viral retinitis 9% (N=106), toxocariasis and toxoplasmosis 7% and 7% respectively (N=83), multiple sclerosis 3% (N=35), tuberculosis 3% (N=35), fungal 3% (N=35), and Birdshot retinochoroidopathy 3% (N=35). Seven percent (N=83) of images were from randomly selected healthy patients. The performance of the model for the detection of vitritis was good with a sensitivity of 96% (95% CI: 0.89-0.98) and a specificity of 87% (95% CI: 0.78-0.89). The area under the ROC curve was 0.92 (95% CI: 0.78-0.99). For the classification of vitritis into 6 classes, the global accuracy of the model was 64% increasing to 85% when accepting a maximum error margin of one class (95% CI: 0.76-0.94).
Conclusions :
We describe a new deep learning model based on UWF fundus imaging that produces an automated and quantitative tool for the detection and grading of vitreous haze with good performance metrics. The model could help implement telemedicine in uveitis clinics and with the evaluation of vitritis in clinical trials by limiting inter-observer reproducibility biases.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.