loading page

Protecting Privacy while Improving Choroid Layer Segmentation in OCT Images: A GAN-based Image Synthesis Approach
  • +6
  • Kiran Kumar Vupparaboina ,
  • Mohammed Nasar Ibrahim ,
  • Amrish Selvam ,
  • Shiva Vaishnavi Kurakula ,
  • Abdul Rasheed Mohammed ,
  • Natasha Mayer ,
  • José Alain Sahel ,
  • Jay Chhablani ,
  • Sandeep Chandra Bollepalli
Kiran Kumar Vupparaboina
Author Profile
Mohammed Nasar Ibrahim
University of Pittsburgh School of Medicine

Corresponding Author:[email protected]

Author Profile
Amrish Selvam
Author Profile
Shiva Vaishnavi Kurakula
Author Profile
Abdul Rasheed Mohammed
Author Profile
Natasha Mayer
Author Profile
José Alain Sahel
Author Profile
Jay Chhablani
Author Profile
Sandeep Chandra Bollepalli
Author Profile

Abstract

To facilitate the training of GAN models for various tasks involving OCT images, we utilized our previously trained residual encoder-decoder (ResUNet) model to generate choroid masks \cite{ibrahim2022volumetric}. Only scans with accurate choroid segmentation, as assessed by an expert grader, were included in the analysis. In particular,  1000 EDI-OCT B-scans from 70 volumes of healthy and AMD subjects as well as 1750 SS-OCT B-scans from 90 volumes of healthy, AMD, and CSCR subjects are graded as good segmentations. Subsequently, the inner (CIB) and outer (COB) boundaries of the choroid, determined based on the choroidal masks, were overlaid onto the B-scans using red (CIB) and blue (COB) color markings. Figure~\ref{fig:DataPrepStep2} illustrates the choroid mask obtained using the ResUNet model, along with the B-scan displaying choroid boundaries marked in red (CIB) and blue (COB). We utilized OCT B-scans and their corresponding choroid boundary-labeled image pairs (CIB-red and COB-blue) as the ground truth. For training the deep learning models, noting the computational complexity, we considered randomly chosen small subsets of 100 EDI-OCT B-scans as well as 100 SS-OCT B-scans. Further, for testing, we considered another 500 EDI and 500 SS-OCT B-scans from the same set of volume scans.