advertisement

---

Editors Selection IGR 16-4

Anatomical Structures: Computer-Assisted Disc Assessment

Comment by Gustavo de Moraes on:

75568 Classification of optic disc shape in glaucoma using machine learning based on quantified ocular parameters, Omodaka K; An G; Tsuda S et al., PLoS ONE, 2017; 12: e0190012

Find related abstracts

---

Optic disc photography is one of the most commonly used methods to document and analyze structural abnormalities due to glaucoma. Nonetheless, it is highly subjective and studies have demonstrated a moderate-to-poor agreement among graders to determine the presence of glaucomatous optic neuropathy and to detect progressive changes. With the advent of artificial intelligence (AI) and more sophisticated neural networks (NN) applied to imaging modalities, it may now be possible to improve the objectivity and repeatability of optic disc photography assessments for glaucoma management.

Omodaka et al. developed an AI-based algorithm for objective classification of the optic disc in patients with open-angle glaucoma (OAG) which was aided by objective, quantitative parameters obtained from ophthalmic examination instruments. In particular, they tested the ability of this algorithm to classify the optic discs of glaucomatous patients into one of the four phenotypes described by Nicolela et al.:¹ focal ischemic (FI), generalized enlargement (GE), myopic (MY), and senile sclerotic (SS). First, three glaucoma specialists performed the classification of optic discs of 163 eyes of 105 OAG patients; cases of disagreement were excluded. Then, they obtained 91 parameters derived from clinical data (n = 7), optic disc topography (n = 22), OCT circumpapillary retinal nerve fiber layer thickness (cpRNFLT) (n = 26), and laser speckle flowgraphy (LSFG) (n = 36). These parameters were used to develop a NN as the machine-learning classifier. The data were divided into training (n = 114 eyes) and validation (n = 49 eyes) sets which were matched from sex, age, visual field mean deviation (MD), spherical equivalent (SE), and intraocular pressure (IOP).

This study may be a useful tool in genetic studies that aim to identify markers associated with different glaucoma phenotypes

The authors found that the NN had an accuracy of 91.2% and Cohen's Kappa of 88%. The most important discriminative characteristics selected by the NN were the SE, age, average nasal rim disc ratio, average cup depth, horizontal disc angle, superior-temporal cpRNFLT, superior-quadrant cpRNFLT, maximum cup depth, and cup area. Disc-type classification by the NN matched the test data at rates of 66.7% for FI, 93.3% for GE, 83.3% for MY, and 100.0% for SS. In some cases, the NN provided ambiguous results due to overlapping phenotypes, which is expected given the subjective nature of the experts' ratings, even after excluding cases of disagreement.

This is an important study as it describes a new technique to classify glaucomatous optic disc phenotypes in a more objective, reproducible fashion. These phenotypes have been shown to have important clinical associations, in particular a predictive value when assessing the risk of progression.^2,3 This new technique may therefore be useful in clinical practice and future studies investigating risk factors for progression. In addition, this may be a useful tool in genetic studies that aim to identify markers associated with different glaucoma phenotypes, such as IOP, disc cupping, and other optic nerve head features.

One study limitation was the small sample size of the testing set given that other AI algorithms described previously in ophthalmology were developed based upon thousands of optic discs or OCT images. NN often require a large number of images in order to extract an adequate number of training features. The fact that the investigators were able to obtain such high accuracy despite its small sample is remarkable. One study strength that may help explain such accuracy was the inclusion of parameters beyond the optic disc photographs (i.e., OCT- and LSFG-derived quantitative measurements) when developing the NN. Future studies using AI in ophthalmology should consider a similar approach in order to optimize their diagnostic performance.

---

• ← Previous
• Next →

---