Abstract
Purpose :
To develop an automated algorithm to reduce noise in 3D optical coherence tomography angiography (OCTA) image volumes.
Methods :
Images were acquired on a OCTA device (AngioPlex, Carl Zeiss Meditec) under an IRB approved protocol from a previously conducted prospective study for secondary analysis. 3x3 mm2 OCT and OCTA volumes with signal strength of 9 or greater were selected from both eyes of healthy subjects without known ocular disease. The automated pipeline was implemented in Matlab (R2022b) and consisted of three phases based on a previously published semi-automated pipeline (Zhang 2020). First, Otsu's method was applied to the OCT volume to perform layer segmentation, and both OCT and OCTA volumes were cropped in the axial direction from the internal limiting membrane to the retinal pigment epithelium layer. Second, curvelet filtering was applied to reduce the speckle noise on the cropped OCTA volume. Third, optimally oriented flux (OOF) was applied to enhance the vasculature on the denoised OCTA volume. The OOF scale levels ranged from 1 to 20 voxels, we used surface-area based normalization and selected the first eigenvalue. The output of the pipeline was subjectively compared to the original images by an expert retinal specialist. An objective comparison of the output to baseline data was made using wavelet noise estimation. Wavelet noise was estimated using the estimate_sigma function from the restoration module of scikit-image Python package.
Results :
A total of 8 images from 4 subjects (50% male, µ=47 years, σ=9 years) were processed without user intervention. The average processing time was 34.1 seconds (cropping ~ 2.1s, curvelet denoising ~ 14.4s, OOF ~ 17.6s). The original images had an average noise of 2.6 +/- 0.8 and denoised images had average noise of 0.006 +/- 0.004, resulting in a 99% reduction in pixel noise (p=1.6x10-5). Visual inspection by an expert reviewer noted better delineation of tertiary vessels in the processed images compared to baseline (Figure 1&2).
Conclusions :
We demonstrate an automated algorithm that significantly decreased the noise in OCTA images while preserving small scale vasculature structure. Future work will continue to improve on metrics such as signal-to-noise ratio, as well as displaying the volumes in three dimensions.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.