Abstract
Purpose :
The choroid, a vascular layer behind the eye's outer retina, crucially supports retinal metabolic functions. Studies connect choroidal structural changes to vision-threatening disorders like age-related macular degeneration (AMD). Accurate identification and quantification of alterations in optical coherence tomography (OCT) images are crucial for informed clinical decisions. In particular, clinicians prioritize choroidal vessel biomarkers for early diagnosis. However, automated detection of choroidal vessels pose a significant challenge due to the intricate structure. This study explores a deep learning approach using generative adversarial networks (GAN) to detect choroidal vessels in OCT images.
Methods :
This is a retrospective study involving 1500 swept-source OCT (SS-OCT) B-scan images of healthy subjects. We adopted a conditional generative adversarial network (GAN), namely Pix2Pix-GAN, which has shown tremendous performance in translating natural images pixel-by-pixel into the target images. Specifically, as depicted in Figure 1, we trained a Pix2Pix GAN model (Generator: Residual U-Net, Discriminator: Patch-GAN) to map OCT images with the corresponding choroid vessel labeled images. We employed 256x256 patches of the OCT image for training to preserve information. The ground-truth labels of the choroid vessels are obtained using our previously validated Phansalkar-thresholding-based method where the segmentations were verified by an expert grader. 1200 images were used for training and 300 for testing. Performance analysis is based on the subjective grading performed comparing the Pix2Pix-GAN-based and ground-truth.
Results :
Figure 2 depicts a qualitative assessment of the choroid vessels obtained by the proposed Pix2Pix-GAN method vis-à-vis ground-truth segmentations, indicating close agreement. The proposed method achieved a mean subjective score of 94.3% against ground-truth.
Conclusions :
The proposed choroid vessel segmentation method based on Pix2Pix GAN demonstrated close agreement with ground truth segmentation. Further studies on diseased scans and other OCT modalities.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.