Deep learning technique outshines AI in detecting glaucoma progression
A new deep learning approach can better discern changes in the eyes of glaucoma patients, according to a new study in Biomedical Optics Express.
Glaucoma is a group of diseases that damage the eye’s optic nerve and can cause vision loss, including blindness. Although there is no cure, early detection and treatment can delay its progression. The progression is marked by complex structural changes in the optic nerve head tissues, such as the thinning of retinal nerve fiber layers and the width of membranes.
Current deep learning methods applied to optical coherence tomography, which uses light to take cross-section images, can detect these changes automatically, but existing methods require a different tissue-specific algorithm to examine each type of tissue. This is also computationally expensive and prone to segmentation errors.
The researchers, from the National University of Singapore and elsewhere, created a new, customized deep learning approach with one algorithm that automatically segments and highlights six different structural parameters of the optic nerve head at the same time.
The approach, called the Dilated-Residual U-Net, or DRUNET, is inspired by U-Net, a convoluted neutral network developed for biomedical image segmentation. DRUNET is comprised of two “towers.” One is a downsampling tower to capture contextual information, such as the spatial arrangement of the tissues; the other is an upsampling tower to capture local information, like tissue texture.
The study authors recruited 100 subjects at Singapore National Eye Center and ended up with 40 healthy controls, 41 individuals with primary open-angle glaucoma, and 19 with primary closed-angle glaucoma. Researchers used manual segmentation to train the algorithm to identify and isolate optic nerve head tissues and employed some data augmentation, since they had a relatively small data set of scans. The scans were split into training and testing data sets.
In testing, overall DRUNET performed significantly better at segmenting and highlighting almost all local and contextual features of the tissues in the tomography images of the optic nerve head than other deep learning methods, researchers found. For one type of tissue, retina pigment epithelium (which lines and protects other tissues), DRUNET’s performance was similar to the other deep learning methods.
DRUNET is also less computationally expensive and faster because it needs fewer trainable parameters. The entire DRUNET network consists of 40,000 trainable parameters. In contrast, the deep learning approach the researchers had been using previously (a patch-based approach) requires 140,000 trainable parameters.
The study authors did acknowledge some limitations to the study. For instance, the accuracy of the algorithm was validated against the manual segmentation provided by only one expert observer. The algorithm was also trained with images from just one machine; it is not known whether it would perform the same way for multiple optic coherence tomography devices.
The researchers hope to extend the use of DRUNET to 3D segmentation.
“With the complex morphological changes occurring in glaucoma, a robust in vivo extraction of these structural parameters could eventually help clinicians in the daily management of glaucoma, thus increasing the current diagnostic power of [optic coherence tomography] in glaucoma,” the study authors say.