Deep Learning Methods Estimate Optic Nerve Deformation With High Accuracy

Deep learning methods, which use image sequences, can estimate optic nerve head deformation with high diagnostic accuracy, according to study findings published in Eye.

Researchers assessed the structural transformation of the optic nerve head region over time in primates from the LSU Experimental Glaucoma Study (LEGS; 12 laser treated eyes and 12 fellow eyes) and individuals from the Diagnostic Innovations in Glaucoma Study (DIGS; ClinicalTrials.gov Identifier: NCT00221897; 36 eyes with glaucomatous progression and 21 longitudinal eyes) using 3 deep learning methods and 2 classical computation methods. The team assessed the biomarker’s diagnostic accuracy using area under the receiver operating characteristic (AUROC) curves.

For the LEGS analysis, AUROC values were 0.83 (95% CI, 0.78-0.88) for all of the deep learning methods, which demonstrated a lower diagnostic accuracy compared with both legacy methods (AUROC, 0.86 and 0.94). The investigators attribute this to image alignment errors in the confocal sequences.

For the DIGS assessment, AUROC values ranged from 0.82 to 0.93 for the deep learning methods compared with 0.86 for both of the classical computational methods. Differences between the deep learning methods were not statistically significant, according to the report.

“Dense [optic nerve head] structural changes estimated using the optical flow methods can provide insights into the changes in the [optic nerve head] architecture and associated structural reorganization in glaucoma,” according to the study authors. “Therefore, biomarkers based on raw measures of underlying structural deformation can not only be useful for detecting glaucoma progression but can also be useful for understanding the characteristics of the [optic nerve head] deformation in patients with
varying risk factors such as age, race and other systemic conditions.”

Study limitations include the use of a limited number of retinal datasets for training the network to achieve high diagnostic accuracy.