Researchers were able to develop a custom hybrid deep convolutional neural network (DCNN), that could accurately and automatically identify the scleral spur on unseen anterior segment optical coherence tomography (AS-OCT) scans, according to a study published in British Journal of Ophthalmology.
They sought to offer an easier method for locating the scleral spur and quantifying structures portrayed on AS-OCT images for the diagnosis and management of primary angle-closure glaucoma (PACG).
After training on a series of 620 images from 42 patients, the researchers included 200 AS-OCT images from 100 eyes of 58 patients. They used a segmentation approach for the DCNNs’ landmark localization and the AS-OCT tissues’ segmentation. They were able to automatically quantify the AS-OCT measurements upon the defining of the scleral spur and the segmentation of the anterior segment intraocular tissues. Investigators used 3 identifiers to locate the scleral spur: a change in curvature in the corneoscleral interface; the posterior end of the trabecular meshwork; and the posterior end of a protruding structure along the cornea and sclera.
In this study, investigators developed a custom hybrid DCNN that they called a full-resolution residual U-Net (FRRUnet), which drew from the benefits of both the U-Net and full-resolution residual network (FRRnet) to locate the scleral spur. U-Net, FRRnet, and the FRRUnet were used together for the segmentation of the AS-OCT structures.
The researchers compared the proposed segmentation approach against a regression approach, which both used DCNNs, to a trained non-expert, a trained medical student and a fellowship-trained glaucoma expert well versed in AS-OCT analysis. The segmentation approach was closer to human observers for all cases.
Intraclass correlation for each observer pair for the X and Y coordinates of the SSL showed that the machine’s scleral spur marking was in high agreement with human graders. In intraobserver tests, the root mean square difference for the machine intraobserver test was significantly smaller than most human intraobserver tests (though less so for the trained non-expert). Machine SSL prediction generally had lower variability than human graders did.
The researchers validated the AS-OCT segmentation performance of the trained network using the Dice coefficient to assess the similarity between the manual segmentation and DCNN segmentation. The DCNN isolated the anterior segment structures with a Dice coefficient of 95.7%.
Interobserver test results demonstrated good to excellent agreement between observers, especially between machine and human. Human-human counterpart results were similar for measurements with lower ICC between machine and human while intraobserver test ICC for machine was higher than human. Limitations of the study included the small size of the dataset.
“This is especially useful for modern swept-source AS-OCT which provides a 360° scan of the eye and as many as 64 cross-section cuts of the anterior chamber angle per eye,” the research explains. “The automated identification of the scleral spur reduces variability of human graders and speeds up image analysis to provide a more comprehensive evaluation of the anterior chamber angle.”
Disclosure: Some study authors declared affiliations with the biotech or pharmaceutical industries. Please see the original reference for a full list of authors’ disclosures.
Pham TH, Devalla SK, Ang A, et al. Deep learning algorithms to isolate and quantify the structures of the anterior segment in optical coherence tomography images. Br J Ophthalmol. Published online September 26, 2020. doi:10.1136/bjophthalmol-2019-315723