Abstract
Purpose :
The quality of optical coherence tomography angiography (OCTA) images is crucial for the accurate interpretation of morphological changes in the retinal vasculature and affects the quantitative analysis results. We demonstrated self-supervised denoising approaches to investigate the repeatability of quantitative parameters, e.g., vessel density (VD) and perfusion density (PD), on OCTA images before and after denoising.
Methods :
Self-supervised denoising with Noise2Void [1] and the adaptation StructNoise2Void [2] were trained to consider spatially correlated noise structure. 3D and 2D U-Nets were used for 3D OCTA volume denoising and 2D slab denoising respectively. Angio 6 × 6 mm scans with 500 A-lines × 500 B-scans from PLEX® Elite 9000 SS-OCT (ZEISS, Dublin, CA) were used for training and evaluation. To train the models, 54 × 4 images from superficial, deep, retina and choriocapillaris slabs were utilized for 2D slab denoising and 54 × 500 B-scans were used for 3D volume denoising. Multi-layer segmentation was performed after 3D volume denoising and the en face OCTA slabs were generated. Automatic threshold, based on the en face images, was used to obtain the binary images for VD and PD measurement. VD and PD were calculated in a 6 mm circle of the retina slab (Figure 1) before and after denoising. 15 healthy subjects and 15 diabetic retinopathy patients with three repeated scans were enrolled. The coefficient of variation (CV) was measured to test the repeatability of the quantitative analysis on retinal slab before and after denoising. A decrease in CV indicates improvement in the repeatability.
Results :
CVPD improved by 19.8% (p=0.035) and 1.7% (p=0.43) after 2D and 3D denoising, respectively (Figure 2). CVVD improved by 2.0% (p=0.45) after 2D denoising but increased by 12.1% (p=0.072) after 3D denoising (no statistical difference before and after denoising).
Conclusions :
The preliminary results from this pilot study showed the self-supervised denoising approaches on both 2D slab and 3D volume may provide improved repeatable measurements in quantifying the vasculature perfusion density. The bias that these denoising models may introduce in the VD and PD measurements were not analyzed here. Future studies will be conducted to investigate the bias and also to investigate the influence of image quality before and after denoising.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.