Segmentation of Drusen in En Face OCT Using Pix2Pix GAN and Embossing

Abstract

Purpose : We propose a novel deep learning approach using a Pix2Pix generative adversarial network (GAN) and image embossing for the segmentation of drusen on en face optical coherence tomography (OCT) images flattened at the Bruch’s membrane (BM) layer.

Methods : In this study, a retrospective dataset of swept-source OCT from patients with dry age-related macular degeneration (AMD) was analyzed. Exclusion criteria were exudative lesions, high myopia, and geographic atrophy. Macula-centered 12x12-mm volumetric scans were acquired using the PLEX Elite 9000 (Carl Zeiss Meditec, Germany). The BM was segmented in normalized and shadow-compensated B-scan using an automated algorithm, and volumetric smoothing was applied during post-processing. The volume was flattened at the BM, and 50 en face slices taken 0 to 100 microns above the BM were averaged to generate an averaged en face image. En face images were cropped at the 6x6 mm center and resized to 256x256 pixels. Image embossing was applied to all images in the four cardinal directions using five pixels of translation. Ground truth of drusen was obtained by manual annotation by an expert clinician. A Pix2Pix generative adversarial network (GAN) model was used to segment the drusen on the embossed en face images. A random 80:20 training-testing data split was used with data augmentation applied to the training set. Dice coefficient (DC) was used to assess performance of the segmentation algorithm.

Results : 25 volumetric OCT scans from 20 patients with dry AMD were used in the study. During data augmentation, training data was expanded to 4000 en face images. Mean testing DC values (± standard deviation) for the Pix2Pix GAN algorithm before and after embossing were 0.702 (± 0.053) and 0.746 (± 0.049), respectively.

Conclusions : Image embossing prior to the use of a Pix2Pix GAN architecture allows for accurate and automated drusen segmentation on en face OCT.

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

 

Image embossing of en face OCT and drusen segmentation using Pix2Pix GAN.