Abstract
Purpose :
To evaluate the diagnostic accuracy of a multi-modality deep learning (DL) classifier using wide optical coherence tomography (OCT) optic nerve head (ONH) cube scans in eyes with and without axial high myopia.
Methods :
This cross-sectional study included 371 primary open-angle glaucoma (POAG) eyes (mean (95% CI) visual field MD -3.78dB (-4.16, -3.41)) and 86 healthy eyes without axial high myopia (axial length (AL) <26 mm) and 92 POAG eyes (mean (95% CI) visual field MD -4.05dB (-4.85, -3.26)) and 44 healthy eyes with axial high myopia (AL≥26mm) from the Diagnostic Innovations in Glaucoma Study (DIGS). We used OCT cube scans of 30deg x 25deg (8.7 x 7.3mm) centered on the ONH, which was formed by 121 horizontal B-scans. The multi-modal DL classifier combines the output features of 3 individual VGG16 models applied on the Spectralis ONH cube scans as follows: 1) texture-based enface image, 2) circumpapillary retinal nerve fiber layer (cpRNFL) thickness map image and 3) scanning laser ophthalmoscope (SLO) image. Area under the receiver operating curves (AUROC) adjusted for both eyes, AL, age, Bruch’s membrane opening area and image quality were used to compare different approaches.
Results :
Adjusted AUROCs were 0.91 (95% CI = 0.87, 0.95) for the multi-modal DL model, and significantly higher (p-value ≤0.05 for all comparisons) than individual VGG16 model: 0.83 (0.79, 0.86) for texture based en-face image, 0.84 (0.81, 0.87) for cpRNFL thickness map, and 0.68 (0.61, 0.74) for SLO image. A subset analysis of high myopic eyes with AL ≥26mm showed significant higher diagnostic accuracy (AUROCs (95% CI)) of multi-modality DL model 0.89 (0.86, 0.92) compared to texture based en-face image 0.83 (0.78, 0.85), cpRNFL 0.85 (0.81, 0.86) and SLO image 0.69 (0.63, 0.76) with p-value ≤ 0.05 for all comparisons.
Conclusions :
Combining the cpRNFL thickness map with texture based en-face images showed higher ability to discriminate between healthy and glaucoma in high myopic eyes than thickness maps alone. While more work is needed, it is likely that texture based en-face images have a role in differentiating high myopic eyes with glaucoma from those without glaucoma.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.