Abstract
Purpose :
To demonstrate automated capillary segmentation in adaptive optics – optical coherence tomography (AO-OCT) images.
Methods :
AO-OCT volumes were acquired from the FDA multimodal AO imager focused on the inner retina. A trained grader generated the vessel plexus projections from the averaged AO-OCT volumes. We trained a UNet-based convolutional neural network to segment retinal capillaries in en face projections of the superficial vascular plexus (SVP), intermediate capillary plexus (ICP), and deep capillary plexus (DCP). The network was trained with random 128x128 pixel patches from 177 automatically contrast-corrected projections of each plexus from 18 eyes in 18 subjects with a 20% validation split. We trained four models: one with all-plexus images and three with plexus-specific images only. The models’ results were compared with a trained grader's manual capillary segmentation. We evaluated segmentation performance based on Dice coefficient (DC), pixel-wise precision, recall, and accuracy on a held-out 20% test split.
Results :
Our all-plexus model achieved good segmentation results with an overall DC of 0.701, precision of 0.686, recall of 0.749, and accuracy of 0.929. All-plexus and plexus-specific model performance had comparable results for each plexus, with slightly better DC in the all-plexus model in SVP and ICP, but worse in DCP projections.
Conclusions :
This is the first application of deep learning techniques for automated vessel segmentation in AO-OCT projections, achieving a good DC, precision, recall, and accuracy and demonstrating the applicability of segmentation techniques to this imaging modality. Difference in results across plexus-specific models may reflect a trade-off in training sample size and task-relevance; however, statistical significance of these results is not determined from this study. Future work is needed to further understand whether automated segmentation should be optimized for each plexus.
This is a 2021 ARVO Annual Meeting abstract.