Deep learning, an advanced form of machine learning, is becoming known for its ability to make sense of data in a quick and automated manner.* This is especially useful in the discovery of glaucoma in patients, as the disease is a leading cause of preventable blindness.

To better understand how deep learning can help the screening, detection, and diagnosis of glaucoma, scientists reviewed existing research on the effectiveness of the technique in combination with known glaucoma diagnostic methods.

What is deep learning?

Most people are familiar with artificial intelligence, which is the simulation of intelligent behavior in computers, as it has been making headlines for the past few years. Machine learning, a form of artificial intelligence, finds patterns in data through the help of humans who identify errors.

Deep learning, a subset of machine learning, is more advanced. Deep learning is an automatic analysis of data that requires no human intervention after initial data input and training are complete. It discovers patterns in data that can be applied in practice.

Deep learning has been used in combination with several glaucoma diagnostic methods, such as:

• Screening with fundus photography

Fundus photography, a low-cost option for screening eye diseases, is often used to detect diabetic retinopathy, which can lead to glaucoma as well as glaucomatous damage. There are mixed reviews on the effectiveness of deep learning in the analysis of fundus photography, as many factors must be perfected before results can be applied in clinical settings. One example is the quality of the fundus imagery, which may render inaccurate results, leading to missed detection or unnecessary referrals.

• Diagnosis with optical coherence tomography

Optical coherence tomography (OCT) is used to detect glaucomatous structural damage through the measurements of the optic nerve head, macula, and the retinal nerve fiber layer. The research review explains that “Deep learning has been shown to improve assessment of damage on raw SDOCT [spectral domain optical coherence tomography] images and visual field data, which could improve the use of these tests in clinical practice.”

• Diagnosis with standard automated perimetry

Deep learning in standard automated perimetry, also known as visual field testing, has performed similarly to, if not better than, human expert graders looking for glaucomatous damage. One study found a limitation, however, in that it was unable to show how much earlier the deep learning technique could detect damage when compared with a conventional technique.

• Diagnosis with structure and function

Researchers are hopeful that deep learning algorithms can make better sense of the vast data derived from structural and functional tests that evaluate the optic nerve and macula. It is likely that this deep learning-derived analysis can be done by clinicians in a manner that is comparable to human experts in diagnosing the disease.

• Glaucoma progression

Only a few published studies have investigated the detection of glaucoma progression. They were able to show the likelihood of progression, but were unable to accurately predict when or how quickly progression might occur.

In essence, though deep learning shows a lot of promise, more rigorous research and testing must be done to improve the sensitivity and specificity of the results gleaned by applying deep learning techniques to the screening, diagnosis and detection of glaucoma.

*Thompson, A. C., Jammal, A. A., & Medeiros, F. A. (2020). A Review of Deep Learning for Screening, Diagnosis, and Detection of Glaucoma Progression. Translational Vision Science & Technology, 9(2), 42. https://doi.org/10.1167/tvst.9.2.42