Fovea-Attentive Residual Network for Extracting 3D Angiograms from Single Volumetric OCT Scan

Optical Coherence Tomography Angiography (OCTA) has drawn numerous attentions in ophthalmology. However, its data acquisition is time-consuming, because it is based on temporal-decorrelation principle thus requires multiple repeated volumetric OCT scans. In this paper, we propose a Fovea-Attentive Residual Network (FAR-Net). By combining an attention mechanism and residual learning, it is able to extract capillary-level angiograms directly from a single OCT scan. The segmentation results of the inner limiting membrane and outer plexiform layers and the central $1\times1$ mm$^2$ field of view of the fovea are employed in the fovea-attention mechanism. So the influences of large retinal vessels and choroidal vasculature on the extraction of capillaries can be minimized during the training of the FAR-Net. The residual learning is used to take advantages of the residual block's shortcut mechanism, so as to avoid the gradient vanishing and accelerate the network convergence. The quantitative and qualitative results demonstrate that the FAR-Net has the capacity to better-visualizing capillaries around the foveal avascular zone than the existing work using a U-Net architecture.