Abstract
Purpose :
Dysfunction of the choroid layer is associated with various posterior segment eye diseases such as age-related macular degeneration (AMD) and central serous chorioretinopathy (CSCR). For accurate disease screening, clinicians seek a quantitative assessment of the choroid layer based on ubiquitous optical coherence tomography (OCT) images. To this end, we attempted a novel image translation deep learning approach to accurately segment the choroid layer using OCT images.
Methods :
This is a retrospective study involving 994 OCT B-scan images of healthy subjects. Motivated by the performance of the Pix2Pix generative adversarial network (GAN) architecture to translate natural images pixel-by-pixel in relation to the target images, we trained a Pix2Pix GAN model with the residual encoder-decoder network as a generator to map OCT images with the corresponding choroid annotated images. Train-test split is 747:247 where test data is blind to training. For training, ground-truth labels of the choroid (inner-boundary: red color, outer-boundary: blue color) are obtained using our previously validated exponentiation method where all are verified by an expert grader. Only images with accurate choroid segmentation are considered as ground-truth. Performance analysis is performed based on the Dice coefficient (DC) between the algorithmic and ground truth segmentations.
Results :
On the 247 test images, the proposed method achieved a mean Dice coefficient of 97.50%. Visual comparison indicated close agreement between the proposed and ground-truth choroid segmentations.
Conclusions :
The proposed choroid layer segmentation method based on Pix2Pix GAN demonstrated close agreement with ground truth segmentation. This study showcases the potential application Pix2Pix GAN in various image segmentation tasks.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.