Protecting Privacy while Improving Choroid Layer Segmentation in OCT
Images: A GAN-based Image Synthesis Approach
Abstract
THIS WORK HAS BEEN SUBMITTED TO THE MACHINE LEARNING: SCIENCE AND
TECHNOLOGY FOR POSSIBLE PUBLICATION. COPYRIGHT MAY BE TRANSFERRED
WITHOUT NOTICE, AFTER WHICH THIS VERSION MAY NO LONGER BE ACCESSIBLE.
The choroid, positioned behind the retina, nourishes the retina by
supplying oxygen and nutrients. Choroidal structural changes are
associated with severe vision-threatening conditions including
age-related macular degeneration (AMD) and central serous
chorioretinopathy (CSCR). Optical Coherence Tomography (OCT) imaging
enables the visualization of choroidal changes, and clinicians rely on
quantifying choroidal biomarkers through segmentation of the choroid
layer in OCT scans for precise diagnosis and disease management.
Accordingly, various attempts are made at automated choroid layer
segmentation, however, their practicality is constrained by the limited
and biased nature of training data. Privacy regulations hinder data
aggregation, and supervised machine learning requires substantial
annotated data. To tackle this, we propose an innovative image synthesis
approach using generative adversarial networks (GANs). It involves a
three-step process: generation of choroid-labeled B-scans using a
standard GAN architecture, the transformation of these scans to
unlabeled B-scans via Pix2Pix-GAN, and the training of a Pix2Pix-GAN
choroid segmentation model using the synthesized data. To demonstrate
the generalizability and efficacy, we evaluated the proposed choroid
segmentation algorithm on the real B-scans from two different OCT
imaging devices: enhanced depth imaging (EDI) and swept-source (SS) OCT,
yielding Dice coefficient values of84.84% and 85.15%, respectively,
buttressing its effective-ness. Further, qualitative performance
analysis, including manual grading, confirms that the synthesized
choroid-labeled images are distinct from real images, thus ensuring data
privacy. The proposed methodology marks an initial step towards
developing a comprehensive choroid layer quantification tool using
synthetic images, and its adaptability makes it versatile for various
medical image segmentation challenges.Â