Abstract
Purpose :
Projection artifact due to vessels within the superficial layers may adversely affect the interpretation of OCT-A. Conventional slab-subtraction-based projection artifact removal (PAR) methods that leverage the anatomic locations of retinal vascular plexuses are generally subject to segmentation errors, making their applications to eyes with retinal vascular diseases more difficult. Here we present a segmentation-independent deep-learning (DL)-based PAR algorithm for OCT-A.
Methods :
A 3D-PAR algorithm inspired by the conventional 2D slab-subtraction method is firstly developed. While capable of removing the projection artifact, this 3D-PAR method yet requires an accurate layer segmentation to create the slabs, and the operation of PAR needs to be performed on the whole 3D volume at once. To overcome these technical limitations, a U-net based DL model is then trained to identify and remove the projection artifacts on a B-frame basis without the input requirement of layer segmentation. OCT-A volumes including healthy and diseased cases are manually inspected for segmentation accuracy and processed to serve as the teacher data. The present U-net based DL model is tested on both healthy eyes and diseased eyes that include the diabetic retinopathy and choroidal neovascularization cases.
Results :
OCTA images of 3×3 mm2 (320×320 pixels) or 6×6 mm2 (512×512 pixels) sizes are acquired with a swept-source OCT (DRI-OCT Triton; Topcon, Tokyo). Fig 1 shows a B-frame of flow signal being overlaid on the structure OCT image of a healthy subject. It can be observed that the false flow signal has been minimized after PAR. A comparison of OCT-A images processed before and after PAR is shown in Fig.2.
Conclusions :
Without the dependency on layer segmentation and the influences from other frames in the volume, the newly developed algorithm is shown to provide a practical advantage to minimize projection artifact in the interpretation of OCT-A.
This is a 2020 ARVO Annual Meeting abstract.