Purchase this article with an account.
Mohamed Abou Shousha, Amr Elsawy, Taher Kamel Eleiwa, Mohamed Tolba, Collin Chase, Eyub Ozcan, Mohamed Abdel-Mottaleb; Artificial Intelligence Algorithm for the Diagnosis of Corneal Graft Rejection. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4310.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To report an artificial intelligence (AI) algorithm for the diagnosis of active corneal grafts rejection (AGR) and rejected corneal grafts (RGR) using corneal optical coherence tomography (OCT) images.
OCT images were obtained from 92 eyes of 82 patients using OCT (Envisu R2210, Bioptigen, Buffalo Grove, IL). Twelve thousands OCT images were randomly selected (3,900 AGR, 3,900 RGR, and 3,900normal controls images) and used to train and validate the AI model. The images were labeled using the diagnosis made by Bascom Palmer Eye Institute cornea specialists. The images were randomly divided into non-overlapped training (80%), validation (10%), and testing (10%) datasets. The AI model was developed using deep convolutional neural networks. The accuracies and the area under receiver operating characteristic curve (AUC) were computed.
The developed AI algorithm was able to achieve overall classification accuracies (i.e., the prediction of AGR, RGR, and normal controls) of 99.46%, 99.05%, and 99.23% in the training, validation, and testing, respectively. Individually, the model achieved accuracies of 100% for AGR, 100% for RGR, and 98.4% for normal controls in the testing. The AUC values were greater than 0.99 for AGR, RGR, and normal controls in training, validation, and testing.
Our AI system trained on OCT images is a novel technique for the autonomous diagnosis of corneal grafts rejection.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only