Purchase this article with an account.
Jingyuan Yang, Kaibin Tian, Weihong Yu, Qijie Wei, Dayong Ding, jianchun zhao, Xirong Li, Youxin Chen; Diagnostic performance of deep learning for multiple retinal diseases based on wide-field fundus photographs of true color. Invest. Ophthalmol. Vis. Sci. 2022;63(7):179 – F0026.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deep learning (DL) technology in recent studies has shown that automated screening, assisted diagnosis, and grading of retinal diseases using conventional fundus photographs (FPs) of about 50° view angle are efficient in saving workforce and time. Compared to conventional FPs, wide-field (WF) images of true color can provide wider view with sufficient resolution, which might further improve the efficiency of screening and DL models, but have not been investigated. Therefore, we investigated the diagnostic performance of DL models for 8 common retinal conditions using fundus photographs captured using Zeiss Clarus 500, by adapting conventional FP models into WF images.
We retrospectively analyzed 474 WF images from 197 consecutive volunteers with retinal conditions, including healthy condition, pathologic myopia, macular abnormalities, retinal vein occlusion, retinal detachment, diabetic retinopathy, after laser photocoagulation treatment, and other abnormalities, which were diagnosed by a panel of retinal specialists. An image could be diagnosed with multiple conditions simultaneously. 258 images without any annotations were used to adjust DL models, and the other 216 images with annotations were used to test the models (Table 1).Considering that WF images show similar appearance in lesions and structures with conventional FPs, we have tried 3 different methods. A model pre-trained on conventional FPs was used to classify whole WF images, and splitted images from WF images, respectively. We also develop a novel method, Global-Local Consistency (GLC), to reach similar diagnostic result between whole and splitted WF images (Figure).
The GLC method was identified as the best diagnostic method for WF images with the highest mean AUC (area under the curve value, 0.8735 vs 0.8541-0.8613). This method was superior to previous models in all conditions except health conditions and retinal vein occlusion (Table 2).
GLC method showed better diagnostic performance for common retinal conditions in this pilot study, even training with a relatively small sample size. Benefit from the similarity between the intrinsic feature of true color of Clarus images and CFPs, model pre-trained on CFPs can be directly used to predict WF images. Our GLC method combining with Clarus facilities might help improve the process of screening retinal diseases.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only