Deep learning-based scatterer density estimation reveals age-related reduction of scatterers in inner retinal layers

Abstract

Purpose : Retinal ganglion cells (GCs) are known to be sensitive to aging and glaucoma. But OCT can only indirectly access the GCs through the layer thicknesses.

We use our developed deep-learning based scatterer density estimator (SDE) to access the retinal aging. Since the scatterer density (ScD) of GC-related layers may indicate the densities of GC, its axon, and synapse, the SDE might be capable of assessing the are-related alteration of GC-related tissues.

Methods : Our SDE is a neural-network based parameter estimator trained by fully numerically simulated OCT speckle image patches (16 x 16 pixels). This was trained by 80,000 images patches (16 x 16 pixels) and estimates a ScD from a patch.

6 × 6 mm² macular OCT images were acquired from 24 eyes of 24 normal subjects [41.2 ± 12.9 y/o (mean ± standard deviation, ranges from 20 to 63 y/o)]. The nerve fiber layer (NFL), GC layer (GCL), and inner plexiform layer (IPL) were segmented by Iowa reference algorithm. The retina was sectorized as shown in Fig. 1. Three more sectors, i.e., the combined-inner and outer rings and whole area of 5.5-mm circle were also used. For each sector-layer combinations (SLCs), the mean ScD and OCT intensity were computed. The correlation between the age and mean ScD and OCT were computed. GC complex (GCC = NFL + GCL + IPL) thickness vs age correlation was also computed.

Results : Fig. 1 shows the mean ScD and OCT intensity, and Fig. 2 summarizes the correlations. The regression lines were shown only if the correlation was significant. Although GCC thickness did not show significant correlation with age, ScD showed moderate negative correlations to the age (r = -0.55, -0.53, -0.54, for NFL, GCL, and IPL) for the whole 5.5-mm sector. The OCT intensity also showed significant but weaker correlations than ScD (r = -0.52, -0.40, and -0.44 for NFL, GCL, and IPL).

The number of significant sectors for ScD vs OCT were 8 vs 7 (ScD vs OCT, for NFL), 3 vs 5 (for GCL), and 9 vs 7 (for IPL). At 19 SCLs, ScD showed higher correlations than OCT (marked with red *), while OCT showed higher correlation only at 3 SLCs (blue *).

Conclusions : ScD of GC-related layers would be a better indicator of age than the OCT signal intensity and the GCC thickness.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

The mean ScD and OCT intensity.

 

Correlations of age vs ScD and OCT intensity.