gNet3D: Fully-3D Deep Learning for Detecting Glaucoma
We are pleased to announce the publication of our latest journal paper in the Special AI issue of Translation Vision Science and Technology (TVST). The paper, entitled A 3D Deep Learning System for Detecting Referable Glaucoma Using Full OCT Macular Cube Scans, details a method for the detection of referable glaucoma from 3D SD-OCT images of the macula.
Glaucoma affects more than 3 million Americans and is the leading cause of irreversible blindness in the U.S. and the world. One of the problems in diagnosing glaucoma is that there is no single test that catches all glaucomatous eyes without generating a large number of false referrals. We applied deep learning to this problem in the form of gNet3D, a fully-3D convolutional neural network (CNN) based on labels derived from glaucoma specialists. To further improve the results, we applied our data homogenization process to bring the input volumes into the same scale, reducing the variance of the inputs and simplifying the resulting network.
The output of the network — referral vs. non-referral — circumvents the need for a normative database, making comparisons directly to the disease state as opposed to using normal ranges. The results were compared against a state-of-the-art 3D OCT-based deep learning framework previously reported in the literature:
gNet3D is capable of very accurately referring patients for glaucoma in a dataset representative of a real-world clinical setting. Studies comparing these results against those of non-specialist ophthalmologists are currently underway.