Purchase this article with an account.
Jamie Shaffer, Luis De Sisternes, Julia Owen, Yelena Bagdasarova, Cecilia S Lee, Warren Lewis, Craig Leong, Roger Goldberg, Mary K Durbin, Aaron Y Lee; Automated plexus differentiation for retinal vessels in OCTA data using deep learning. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2139.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
OCTA data is able to differentiate vasculature from different retinal plexuses. However, the divisions between superficial, deep, and avascular regions are typically based on segmentation of structural layers which may not perfectly correspond to the vascular plexus divisions. In this work we introduce an automated method to divide the vasculature observed in OCTA volumes according to their plexus using only the angiographic (not structural) data.
Image slabs from 235 OCTA cubes (33 patients) were handcrafted to contain only vascular data belonging to the superficial, deep, or avascular plexus. In addition to this labeled data, thin slabs from the volume were generated as unlabeled data for qualitative model evaluation. Patient-level partitioning was used to separate images into training, validation, and test sets. Two UNet models were trained and validated (Fig 1). Model 1 used single-class images with image augmentation (flip, rotate, contrast). Model 2 added single-patient synthetic images created by blending a randomly placed and sized region from an adjacent class into each image. Models were evaluated on the reserved test set: quantitatively on both single-class slab and 2-class synthetic images and qualitatively on the unlabeled images.
Both models performed well on labeling single-class images (DICE ≥ 0.90) but less well on 2-class images, particularly when distinguishing deep and avascular regions (Fig 2A). Model 2 performed only marginally better than Model 1, indicating that deep learning algorithms are robust enough to segment multi-class images even when trained only on single class images (Model 1). Predictions on unlabeled slabs were qualitatively evaluated (Fig 2B).
We present a deep learning model that can learn from single-class images and synthetic combinations of those images to annotate the pixels of unlabelled OCTA slabs according to their location within the superficial, deep or avascular plexus. The method looks promising for differentiating the location of different vascular plexuses within an OCTA volume independently from retinal layer structural information.
This is a 2021 ARVO Annual Meeting abstract.
Overview of data preparation and model training.
A: DICE scores for slab single-class and synthetic 2-class images. B: Predictions for superficial, deep, and avascular on unlabeled slab images.
This PDF is available to Subscribers Only