The eyes are sometimes referred to as a window to one’s soul, a phrase with unclear origin, but filled with truth. In fact, our eyes provide a literal window into our brains. This is because neurodegenerative conditions such as Parkinson’s disease and Alzheimer’s disease, as well as eye diseases such as glaucoma, can directly affect a patient’s optic nerve, retinal cells and surrounding visual structures. The eye, or more specifically, the retina, is often known as an extension to our central nervous system (CNS).
One type of cell, the retinal ganglion cell (GC), is a neuronal cell whose axons form the optic nerve. These cells are key to one’s vision as they process and relay optical information to the brain. In neurodegenerative diseases, GCs often degenerate and disappear, eventually leading to blindness. For this reason, being able to quickly identify the extent of such degeneration has proven to be an important biomarker for both diagnosis and treatment monitoring of neurodegenerative diseases. In a study published in Optica, researchers in the US present a new method to visualize and quantify individual GCs.
A weakly supervised segmentation network…
Ophthalmic imaging systems, such as optical coherence tomography (OCT), are clinically used to visualize the different layers of eye tissue, including the GC layer, in order to diagnose and track the progression of eye diseases. However, as these traditional methods are limited to low resolution, they can only measure the thickness of these layers, without providing much information on individual cells.
A state-of-the-art technology, called adaptive optics OCT (AO-OCT), is sensitive enough to image individual retinal GCs. However, the current standard approach for quantification involves manual marking of each AO-OCT volume, which is not only subjective, but also time consuming and impractical for large datasets.
To solve this problem a group of researchers from Duke University, Indiana University, the University of Maryland and the FDA Centre for Devices and Radiological Health has developed an automatic technique for rapid and objective quantification of GC somas (cell bodies). They call their method, which is based on weakly supervised deep learning, “WeakGCSeg”.
This method consists of a three-step process. First, the group uses an automatic algorithm to pre-process an entire AO-OCT volume, to extract the retinal layer containing the GC somas. Second, the researchers pass this volume through a localization network that is trained to produce a probability map indicating the locations of potential somas. Interestingly, the network is not trained using ground-truth segmentation maps, but rather using “weakly” annotated labels in the form of small spheres (with a radius of 2 μm) at each manually annotated cell location. In the final step, the researchers perform post-processing to transform the network’s predictions into individually segmented somas.
… accurately localizes individual ganglion cells
To test their method, the researchers analysed both healthy subjects and glaucoma patients, and showed that their proposed framework was able to accurately segment GC somas of both groups. By using WeakGCSeg’s output, they could differentiate between the two cohorts based on the number and size of the predicted GCs. Not only that, but this framework was able to achieve high detection performance regardless of the imaging device used, while being on par with or exceeding the performance of human expert graders.
In the future, the group hopes to extend this work using larger datasets and looking at different retinal diseases to fully capture the framework’s generalizability. The researchers note that WeakGCSeg also has the potential to be used with other neurodegenerative disorders, such as Alzheimer’s or Parkinson’s disease.
