Abstract
Purpose :
Glaucomatous vision loss may be preceded by enlargement of the cup-to-disc ratio. We propose to validate an artificial intelligence (AI) based cup-to-disc ratio (CDR) grading system that may aid in effective glaucoma-suspect screening.
Methods :
A total of 1500 color fundus images that included the disc were selected randomly from AREDS and SiMES, and all the images from RIMONE1 dataset, an ophthalmic reference image database specifically designed for glaucoma analysis. These images were processed to get local color averages (LCA) and subtracting them to eliminate light intensity or other input error causing from various conditions. A proprietary algorithm for cropping the optic disc area was used to automatically obtain the optic disc images from the full fundus images. These cropped images from the original RGB image set and the LCA set form the final input for this experiment. Proprietary semi-automated software was first implemented by experts to establish the ground truth CDR in each image. Then, deep learning architectures were trained and validated to produce completely automatic outputs. All graded images (1500) were grouped into two categories: CDR <= 0.5 (1056 images) and CDR above 0.5 (444 images) in the two-class model. Additionally, the images were regrouped into three categories for the three-class model: CDR less than 0.5 (1056 images), 0.5 to 0.6 (324 images), and over 0.6 (164 images) with both groups containing a training set of 1200 images and a test set of 300 images. Sensitivity, specificity, accuracy and kappa scores were calculated to evaluate the models.
Results :
The two-class glaucoma model (CDR <= 0.5 and above 0.5) achieved an accuracy of 89.67% (95% CI - 85.65% to 92.87%) with a sensitivity of 83.33% (95% CI - 75.44% to 89.51%) and a specificity of 93.89% (95% CI - 89.33% to 96.91%). The three-class glaucoma model (CDR <0.5, 0.5 to 0.6, >0.6) achieved an overall accuracy of 81.33% (95% CI - 71.2% to 89%, with 243 out of 300 correctly identified) with a high weighted kappa of 0.77 (95% CI - 0.74 to 0.80).
Conclusions :
We have demonstrated a highly accurate and fully automated glaucoma screening system that may be effective for the identification of glaucoma suspects. A further prospective trial can determine the feasibility of the system in clinical settings.
This is a 2020 Imaging in the Eye Conference abstract.