Abstract
Purpose :
Early detection is the key to prevent glaucoma blindness, but the diagnostic rate of glaucoma remains low. We developed a deep learning model for automated glaucoma detection to explore the potential of automated teleglaucoma screening.
Methods :
Fundus color images of eyes with glaucomatous optic neuropathy or normal disc were retrospectively collected from the image database of Taipei Veterans General Hospital after approval by our institutional review board. Based on these image datasets, a deep learning model for automated glaucoma detection was developed based on deep convolutional neural network (CNN). Briefly, fundus images were first scaled and pre-processed to enhance disc region and vessel contour. Region of interest, the optic nerve head, was then identified and marked for further modeling. During the mapping model stage, the images were divided in to 3 groups, training, validation, and test datasets. After ensemble learning stage, the model assigned a probability score ranged from 0 to 1 indicate the possibility of glaucoma.
Results :
Glaucomatous and healthy fundus images (four hundred each) were collected. With the CNN model, the sensitivity and specificity of the classification model in glaucoma diagnosis was 91.8% and 93.0%, respectively with validation dataset (50 images in each group); and 83.3% and 90.9%, respectively with test dataset (25 images each). If public database, including High-Resolution Fundus Image database and Drishti-GS dataset were included, the sensitivity and specificity were 71.6% and 88.7%, respectively.
Conclusions :
With limited image database, the CNN model achieved comparable diagnostic accuracy of glaucoma as existing strategies based on optic disc or optic cup extraction. Further optimization of the model with larger data sets will be investigated to explore the potential of automated glaucoma detection in clinical setting.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.