Deep Learning Provides Insight Into Visual Function in Glaucoma
Ophthalmology, November 2021
Christopher et al. developed combination deep learning (DL) models that improve the estimates of visual function obtained from macula-centered spectral-domain OCT (SD-OCT). They found that these models may help physicians devise appropriate treatment plans for their patients with glaucoma.
The researchers set out by training DL models on thickness maps from SD-OCT macular images to estimate 10-2 and 24-2 visual field (VF) mean deviation (MD) and pattern standard deviation (PSD). They trained separate and combined DL models using data for six layers: retinal nerve fiber layer, ganglion cell layer, inner plexiform layer (IPL), ganglion cell-IPL (GCIPL), ganglion cell complex, and the retina itself. Combination models that incorporated data for all six layers were constructed using linear regression. Each model was equipped only with thickness maps or mean measurements derived from SD-OCT, not with labels of disease status or severity.
Findings for these models were compared with VF measurements. SD-OCT images of each study eye were paired with 10-2 and 24-2 VF findings. Primary outcome measures were R2 and mean absolute error.
Altogether, the researchers assessed 2,408 SD-OCT/10-2 VF pairs and 2,999 SD-OCT/24-2 VF pairs. Images and data were obtained for healthy participants and for patients with suspected or confirmed glaucoma. The 10-2 cohort included 1,051 eyes (563 subjects), and the 24-2 cohort included 1,205 eyes (641 subjects); 1,037 eyes (560 subjects) were common to both cohorts.
The combined DL models estimating 10-2 achieved R2 of 0.82 for MD and 0.69 for PSD; mean absolute error was 1.9 dB for MD and 1.5 dB for PSD. These findings were significantly more accurate than thickness estimates for 10-2 MD (0.61 R2, 3.0 dB) and 10-2 PSD (0.46 R2, 2.3 dB). The combined DL models estimating 24-2 achieved R2 of 0.79 for MD and 0.68 for PSD; mean absolute error was 2.1 dB for MD and 1.5 dB for PSD. The DL models outperformed mean thickness estimates for 24-2 MD (0.41 R2, 3.4 dB) and 24-2 PSD (0.38 R2, 2.4 dB). Among the individual models, data were most accurate for GCIPL (0.79 R2) and ganglion cell complex (0.75 R2).
The authors concluded that DL models provide good estimates of functional status. They emphasized that SD-OCT images contain abundant information that has yet to be tapped; adding DL to the mix could lead to new biomarkers and better personalized treatment plans.
The original article can be found here.