Automated choroid layer segmentation based on wide-field SS-OCT images using deep residual encoder-decoder architecture

Abstract

Purpose : The choroid is a vascular structure of the ocular posterior segment which serves a number of crucial metabolic and homeostatic functions. Choroidal dysfunction is implicated causally in several macular diseases, so there is considerable interest in characterizing choroidal structure in terms of specific biomarkers, using in vivo imaging. However, clinical studies have so far been based on optical coherence tomography (OCT) images corresponding to 6mm × 6mm macular scanning area. A wide-field swept-source OCT (SS-OCT) imaging system provides denser raster scans over a larger, 12mm × 12mm scanning area that includes the optic nerve. The aim of this study was to achieve automated segmentation of the choroid in wide-field SS-OCT scans. A deep learning architecture based on residual encoder-decoder modules proved efficacious for this purpose.

Methods : A retrospective dataset of 613 wide-field SS-OCT B-scans with a resolution of 12mm × 3mm (1024×1536 in pixels) taken from healthy eyes was used in this experiment. Images were captured using a wide-field SS-OCT device (Carl Zeiss Plex Elite 9000). Inspired by the ability of deep residual networks to propagate the information without loss and the success of encoder-decoder architecture in image segmentation, we adopted residual U-Net (ResUnet) for choroidal segmentation (Fig. 1). A random 70:30 data split was considered for training and testing (unseen). Ground truth annotations (masks) required for training the model were obtained by a previously validated segmentation tool and all annotations were verified by clinicians. Images were resized to 256x256. Further, the model was trained for 80 epochs, with a batch size of 8, to minimize Dice coefficient (DC) loss. The DC criterion was also used for performance evaluation.

Results : Training and Testing DC values were found to be 99.42% and 97.81%, respectively, which were desirably high. Segmentation results on representative test images, depicted in Fig.1, visually demonstrate the efficacy of the ResUnet architecture.

Conclusions : The proposed ResUnet approach demonstrated performance close to the ground truth. We next plan to extend the methodology to diseased datasets. Further, we envisage extending the 2D approach to train directly on 3D volumes.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig. 1. Proposed choroid segmentation in wide-field SS-OCT images using ResUnet architecture.