Deep Learning Models Predict Visual Function from Macula Thickness Map

Abstract

Purpose : To apply deep learning approaches to predict visual field (VF) mean deviation (MD) from multiple thickness maps derived from macular imaging.

Methods : Imaging and VF measurements were collected from a cohort of 394 healthy, suspect, and glaucoma participants and 703 eyes. The imaging consisted of 1448 Spectralis horizontal posterior pole (61 b-scans, 30°x25°) images. VF testing in the form of 10-2 (mean ± standard deviation MD = -3.9 ± 6.3 dB) and 24-2 (mean ± standard deviation MD = -4.1 ± 6.3 dB) tests was completed within 180 days of imaging. Participants were randomly divided into independent training (85%), validation (5%), and test (10%) sets. Using standard instrument segmentation, retinal nerve fiber layer (RNFL), ganglion cell layer (GCL), and inner plexiform layer (IPL), ganglion cell – inner plexiform layer combined (GCIPL), ganglion cell complex (GCC), and whole retina (RET) thickness maps were generated Individual deep learning models were trained to predict both 10-2 and 24-2 MD from each layer. Combined deep learning models that use all layers simultaneously were also constructed. The Resnet50 architecture was used and transfer learning was applied by pre-training on a general image dataset (ImageNet). Models were evaluated using R² to compare to predicted and true MD. As a baseline comparison, linear regression models that predicted MD from mean layer thicknesses were also evaluated.

Results : In predicting 10-2 MD, the individual deep learning models, the GCIPL (R² = 0.84), GCL (R² = 0.78), and IPL (R² = 0.74) outperformed other layers. The combined deep learning model achieved an R² = 0.87 (p < 0.001), outperforming mean layer thickness R² = 0.60 (p < 0.001). In predicting 24-2 MD, the combined deep learning model achieved an R² = 0.82 (p < 0.001) while the mean layer thickness prediction had an R² = 0.62 (p < 0.001). Here, the RNFL (R² = 0.73), GCIPL (R² = 0.71), and IPL (R² = 0.70) outperformed other layers. The full results are summarized in Table 1.

Conclusions : Applying deep learning models to macular structural measurements can predict both central and peripheral visual field loss in glaucoma with a high degree of accuracy. The results suggest that deep learning models can provide useful clinical information without relying on time-consuming and subjective visual field testing.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Table 1: Performance of macula layer thicknesses and deep learning models to predict MD of 10-2 and 24-2 visual fields.