Purchase this article with an account.
Tong Li, Ningjiang Chen, Yaofeng Wen, Xiang Xu, Cong Ye, Shaodan Zhang, DaHong Qian, Yuanbo Liang; Differentiation of primary open angle-closure glaucoma and primary open angle glaucoma based on disc image with a deep learning method. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4553.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
POAG (Primary Open Angle Glaucoma) and PACG (Primary Angle Closure Glaucoma) are two major subtypes of glaucoma with different mechanism of optic nerve damage. Usually people use a gonioscope or an AS (Anterior Segment) OCT to measure the degree of angle closure so that they can distinguish between these two subtypes. However, for the scenario of screening at primary care it needs an additional device and test. Furthermore, gonioscope is not easy to use for the primary care. At the same time, fundus camera gradually becomes a popular device for screening eye diseases. If we can distinguish POAG from PACG by only using fundus images, it will make the glaucoma screening more productive.We try to train a deep learning model to classify these two types of glaucoma and highlight the important area for the model’s classification decision. These highlight areas may account for the discrepancy between the two subtypes of the glaucoma on fundus images.
We analyzes 603 patients and 4057 fundus images from theWenZhou Medical University and split into training dataset and evaluation dataset. A model based on ResNet50 was trained to distinguish POAG from PACG. Then we apply a method of model interpretation named GradCam to visualize the hidden feature maps and reflect into the original images.
We get the initial classification accuracy of 0.7253, the area under ROC curve of 0.8071, and the F1 score of 0.7449. The saliency area on the images will help us to check whether the difference between POAG and PACG in anatomy exists or not.
Our method shows a promising capacity to classify POAG and PACG fundus image based on a deep learning method. With more training samples and excluding early-stage glaucoma samples, the classification accuracy will be improved.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only