Abstract
Purpose :
To compare global and sector wise performance of deep learning based algorithm versus Cirrus Optical Coherence Tomography(OCT) for Macula and Optic Nerve Head (ONH) parametres in glaucoma versus normal population
Methods :
We retrospectively looked at Glaucoma and Normal populations at Stanford Univ.Dept of Ophthalmology, identified by a specialist based on disc changes and visual field defects with borderline cases arbitrated by a second specialist.Analysis of ONH and Macula was done with Cirrus software(Carl Zeiss Inc.,Dublin,CA)and a hybrid deep learning approach using Orion(Voxeleron,Pleasanton,CA) and was compared. Orion, being assessed for first time on clinical data, uses deep learning method to delineate ONH from enface view followed by conventional segmentation of ONH and Macula. For ONH-Retinal Nerve Fibre Layer (RNFL) analysis we had n =139 for Glaucoma and n=163 eyes for Normals with RNFL thickness reported peripapillary to the center of ONH. Overlapping the previous set of eyes we had for Macula-Ganglion Cell Inner Plexiform(GCIPL) analysis n=135 for Glaucoma and n=160 eyes for Normals with average thicknesses reported in same elliptical annulus.Receiver operating characteristic(ROC) curves were created and area under curves(AUCs)for each parameter used to gauge performance
Results :
AUCs for GCIPL parametres; Average,Temporal-Superior, Superior, Nasal-Superior, Nasal-Inferior, Inferior, Temporal-Inferior sectors for Cirrus were 0.86,0.83,0.8,0.75,0.81,0.86 ,0.9 and for Orion were 0.85, 0.81, 0.8, 0.72, 0.78, 0.86, 0.9 respectively. AUCs for RNFL parameters; Average,Temporal, Superior, Nasal, Inferior sectors for Cirrus were 0.91, 0.69, 0.85, 0.75, 0.93 and for Orion were 0.9, 0.71, 0.84, 0.79, 0.92 respectively. For GCIPL, best AUC was 0.9;for both Orion and Cirrus; both for temporal-inferior sector.For RNFL,Cirrus had highest performance- AUC of 0.93 followed by Orion-AUC of 0.92; both for inferior sector
Conclusions :
RNFL had marginally better diagnostic performance in this population.Both found most diagnostic information to be in inferior regions of macula and ONH. Voxeleron's algorithm trained on small data set establishes similar diagnostic performance as Cirrus and prospectiveness of using hybrid deep learning in identifying glaucoma cases.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.