Glaucoma patient screening from online retinal fundus images via Artificial Intelligence

Abstract

Objectives: To design and evaluate a novel automated glaucoma classifier from retinal fundus images. <BR> Methods: We designed a novel Artificial Intelligence (AI) automated tool to detect glaucoma from retinal fundus images. We then downloaded publicly available retinal fundus image datasets containing healthy patients and images with verified glaucoma labels. Two thirds of the images were used to train the classifier. The remaining third of the images was used to create several cross-validation evaluation sets with a realistic glaucoma prevalence, to evaluate the classifier's performance in a screening scenario. <BR> Results: 10,658 retinal fundus images from seven different sources were found and downloaded. They were randomly divided into 7,106 for training and 3,551 for validation. Glaucoma prevalence was 24%. Using the validation set, we created 50 random sets of 1,000 images with a 5% glaucoma prevalence. On these sets, the classifier reached a detection rate of 84.1% (CI 95=+-1.1%) and 95.8% (CI 95=+-0.2%) specificity. <BR> Conclusions: A novel glaucoma classifier from retinal fundus images showed promising results as screening tool on a large cohort of patients. A large clinical study is needed to confirm these results.

Deep Learning Application to Detect Glaucoma with a Mixed Training Approach: Public Database and Expert-Labeled Glaucoma Population

“…Only images labeled as glaucomatous or healthy eyes were used from each dataset, and images labeled with other ocular pathologies were excluded. The disease stage in these databases has not been considered since many of them did not report the stage [14].…”

“…To achieve this objective, the study was divided into two different phases. First, a classifier model trained by using images from public datasets [14] was tested on real practice images (initial sample) from the dataset from the Instituto de Oftalmobiologia Aplicada (IOBA) of the University of Valladolid, Valladolid, Spain. Second, the model was retrained with new images from the same IOBA dataset (second sample) and then retested with the same test images (initial sample).…”