Choroid layer segmentation based on synthetic optical coherence tomography marked-unmarked image pairs: A generative adversarial network approach

Abstract

Purpose : Changes in choroidal structure are linked to severe vision-threatening conditions, including central serous chorioretinopathy (CSCR). Optical coherence tomography (OCT) captures choroidal changes, and clinicians seek precise OCT choroidal biomarkers, like thickness, for effective disease management. Despite efforts in automated choroid layer segmentation, practicality is limited due to biased training data and privacy constraints. To address this, our study proposes an image synthesis approach for choroid layer segmentation, relying solely on synthetic choroid-marked and unmarked image pairs, reducing dependency on real OCT databases.

Methods : This retrospective study involved 100 enchance depth imaging (EDI) OCT B-scans, employing an innovative image synthesis method with generative adversarial networks (GANs) (see Figure 1(a)). The proposed three-step methodology involved generating choroid boundary-marked B-scans using a standard GAN, transforming synthetic choroid-marked scans to unmarked B-scans with a conditional GAN (Pix2Pix-GAN), and training a Pix2Pix-GAN choroid segmentation model with synthesized marked-unmarked image pairs. Performance analysis included subjective grading to differentiate synthesized from real OCT scans and calculating Dice coefficient (DC) between algorithmic and ground truth segmentations on real OCT data.

Results : Figure 1(b) showcases stepwise results on representative OCT B-scans, highlighting the effectiveness of our proposed approach. The subjective grading score for synthesis quality is 94%, and the synthetic-image-based segmentation model achieved an accuracy of 84.84% Dice coefficient when tested on real OCT images.

Conclusions : The choroid segmentations from our proposed method align closely with ground truth segmentation. Qualitative analysis, including manual grading, affirms the distinctiveness of synthesized choroid-labeled images, ensuring data privacy. This methodology represents an initial step towards a versatile choroid layer quantification tool using synthetic images, adaptable to diverse medical image segmentation challenges.

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

 

Figure 1. (a) Schematic of proposed methodology; and (b) Stepwise results. Note that Step-3 depicts segmentation performed on real images using the segmentation model trained on synthetic image pairs. Notation: CIB—choroid inner boundary, and COB—choroid outer boundary.