A recent study in Biomedical Optics Express revealed that a new deep learning method could better distinguish changes, particularly in the eyes of those suffering from glaucoma. In essence, glaucoma entails a group of ailments that can compromise the eye’s optic nerve leading to vision loss.
Even though this condition lacks a cure, early detection and treatment can aid in delaying its progression significantly. Glaucoma progression is normally characterized by complex structural alterations that occur in the optic nerve head tissues, for instance, the thinning of the width of membranes and the retinal nerve fiber layers.
Today’s deep learning techniques that are applied to optical coherence tomography, which utilizes light in a bid to take cross-section images, can identify the structural changes automatically. However, these existing approaches need a different tissue-specific algorithm for examining each type of tissue. Furthermore, they are not only computationally costly but also susceptible to segmentation errors.
A team of researchers drawn from National University of Singapore among other places developed a new, modified deep learning method with a single algorithm that automatically splits and selects six different structural parameters relating to the optic nerve head simultaneously.
Dubbed the Dilated-Residual U-Net or otherwise known as DRUNET, is the newly developed approach that is inspired by U-Net, which is a convoluted neural network (CNN) that was designed for biomedical image segmentation. DRUNET is made up of two towers whereby one entails a downsampling tower intended for capturing contextual information while the other is an upsampling tower that is helpful in capturing local information such as tissue texture.
The authors of the study drew together 100 subjects at Singapore National Eye Center. The result of the recruitment was 41 people with primary open-angle glaucoma, 40 healthy controls and 19 individuals with primary closed-angle glaucoma. On the other hand, researchers utilized manual segmentation in training the algorithm to spot and separate optic nerve head tissue. They also applied some data augmentation, as they had a small dataset of scans that were divided into testing and training data sets.
When it came to testing, researchers found out that DRUNET performed considerably better than other deep learning techniques at separating and highlighting almost all contextual and local attributes of the tissues in the tomography images, particularly those of the optic nerve head. However, for the retinal pigment epithelium, DRUNET performed similarly to the other existing deep learning approaches.
Since it requires fewer trainable parameters, DRUNET is less computationally costly and faster. In fact, its whole network boasts about 40,000 trainable parameters. On the contrary, the deep learning method that the researchers were using previously needs 140,000 trainable parameters.
Aside from the successes of the study, the authors also owned up to some limitations. For example, they said that the algorithm’s accuracy was authenticated against the manual segmentation, which was offered by only one professional observer. According to the authors, the algorithm was trained using images from a single machine.
One study author said that a strong in vivo extraction of such structural parameters could ultimately assist clinicians in the everyday management of glaucoma. Doing so would help to boost the current diagnostic capability of optic coherence tomography in the condition.