Towards determining the location of the Preferred Retinal Locus of patients with macular disease: A deep learning-based simulation

Poster Presentation: Tuesday, May 21, 2024, 2:45 – 6:45 pm, Pavilion
Session: Spatial Vision: Machine learning, neural networks

Rijul Saurabh Soans1, Dongcheng He1, Susana T L Chung1; 1University of California, Berkeley

In the absence of a healthy fovea, individuals with macular disease often rely on an alternate retinal location, the preferred retinal locus (PRL), for seeing. However, the factors determining the PRL location are still unclear. Previous studies have reported that the PRL location is not driven by exclusively optimizing visual acuity (Bernard & Chung, 2018) or sensitivity (Chung et al., 2023). This study tests the feasibility of applying the semantic segmentation approach to predict the PRL location based on the eccentricity effect and how visual acuity and sensitivity vary across the retina. We first generated 200 scotoma patterns based on the actual scotoma shape and size of 21 eyes with macular disease (by sampling along each of 360° meridians the location of the edge of each generated scotoma). We then placed each scotoma pattern on the retinal images of 79 healthy eyes to generate 15,800 images. Each sample of our simulated data was a 3-channel vector (256×256). The first channel consisted of the image in grayscale pixel values. The second and third channels consisted of visual acuity and sensitivity values, respectively, generated based on values reported in the literature. For training purposes (using a DeepLabV3 network with 13,000 samples), we defined the PRL region of each vector by imposing certain constraints: it is close to the edge of the scotoma (Chung, 2012); its acuity and sensitivity values are near expected thresholds; and it is located closest to the fovea after satisfying the preceding constraints. Testing with the other 2800 samples showed that between the predicted and “true” PRLs, the mean Jaccard index (percent-overlap of area) was 53.7%±25.7%, and the mean Euclidean distance between their centers was 8.6±22.6 pixels. Our results suggest that a model combining eccentricity and visual performance variation across the retina is feasible in predicting the PRL location.

Acknowledgements: Authors RSS & DH contributed equally. Grant Support: NIH Grant EY030253