Abstract
Purpose :
To develop and validate a deep convolutional neural network (CNN) to robustly segment geographic atrophy (GA) area on optical coherence tomography (OCT).
Methods :
We obtained 6x6-mm OCT scans from the central macular region of either one or both eyes of 134 participants enrolled in a clinical study on age-related macular degeneration (AMD). The scans were acquired using a 120-kHz commercial OCT system (SOLIX; Optove/Visionix, Inc.). A two-stage deep convolutional neural network (CNN) was developed to perform segmentation of geographic atrophy (GA), as illustrated in Fig. 1. In the first stage, a CNN segmented the vascular tissue into the retina and choroid, using Bruch's membrane as the boundary (Fig. 1 A-C). Subsequently, two en face images – a whole OCT volume projection image (Fig. 1 D) and a color-inverted choroidal projection image (Fig. 1 E) – are generated by applying mean projection techniques within the entire OCT volume and the choroidal slab, respectively. In the second stage, the CNN (Fig. 1 F) takes the en face images as input and produces the segmentation result for the GA area (Fig. 1 G). The ground truths for these networks were manually graded by experienced retinal experts. To assess the performance of our method across the entire dataset, a six-fold cross-validation was employed.
Results :
In total 134 participants were included in this study, including 96 participants with dry-AMD, and 38 healthy controls. With the six-fold cross-validation, our proposed method demonstrated notable accuracy in GA area segmentation (Dice coefficient 0.87±0.07).
Conclusions :
A deep learning-based method can accurately segment geographic atrophy in dry-AMD eyes on OCT scans.
This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.