Abstract
Purpose :
Segmentation of en face retinal vasculature using volumetric optical coherence tomographic angiography (OCTA) usually relies on prior retinal layer segmentation, which is time-consuming and error-susceptible. In this study, we propose a deep-learning-based method to segment vessels in the superficial vascular complex (SVC), intermediate capillary plexus (ICP) and deep capillary plexus (DCP) directly from OCTA data volumes.
Methods :
A total of 88 2×2-mm OCT (Fig. 1 A) and OCTA (Fig. 1 B) volumes were acquired on one eye each from 10 brown norway rats using a 50-kHz prototype visible-light OCT system. A deep convolutional neural network (CNN) (Fig. 1 C-E) was designed to segment the vasculature in the SVC, ICP and DCP. The input data are the structural OCT and the corresponding OCTA volumes. The first CNN branch (Fig. 1 C) is three-dimensional and used for segmenting retinal slabs. Retinal slabs from each volumetric OCT were manually delineated by certified graders to establish the ground truth. A projection layer (Fig. 1 D) was designed to generate three capillary angiograms by projecting the maximum flow signal within each corresponding retinal slab. For each angiogram, we used a CNN branch (Fig. 1 E1-E3) to perform binary classification of the flow signal to generate the final outputs (Fig. 1 F-H). The ground truth for the vasculatures within three capillary angiograms were also manually delineated by certified graders. We used categorical cross-entropy as the loss function and Adam optimizer in training.
Results :
66 volumes were used in training, with 22 volumes reserved for testing. Compared to the manually delineated ground truth (Fig. 2), our method achieved high accuracy in the SVC (Dice coefficient = 0.90 ±0.09, mean ± standard deviation) and DCP (0.79±0.10). Accuracy was lower in the ICP (0.55±0.16) due to manual delineation errors (Fig. 2 B2). As an intermediate result in the network reasoning, the retinal slab segmentations showed high overall accuracy (Dice coefficient = 0.93 ± 0.06), indicating that the vessels segmented by our approach appeared to be in the correct plexus or complex.
Conclusions :
Our deep-learning-based method can perform accurate capillary-scale vascular segmentation in the SVC, ICP and DCP. This method provides an end-to-end pipeline from raw volumetric OCTA to en face vascular segmentation without requiring human inputs or manual corrections.
This is a 2020 Imaging in the Eye Conference abstract.