Abstract
Purpose :
Because of the low transverse sampling, current wide-field OCT angiography (OCTA) has a poor lateral resolution thus causes the inaccurate observation and quantification of vascular biomarkers. We propose a deep-learning-based approach for the high-resolution reconstruction of wide-field OCTA, which could be directly applied to the data from commercial OCT systems. It achieves capillary-level retinal vasculature over a large field of view (FOV) and matches the resolution of the angiograms using high transverse sampling.
Methods :
We propose to learn the characteristics of high-resolution capillaries and micro-vessels by only leveraging the information provided by the 3×3 mm2 FOV centered on fovea, and teach them to 8×8 mm2 FOV. Our baseline method is cycle-consistent adversarial learning. However, we found it not only learned the resolution-related features but also mapped the spatial distribution of retinal vasculature, which is undesirable in this application. So we propose to split both the source and target domain OCTA images into 1×1 mm2 patches before feeding them into the adversarial network. We found this approach could efficiently eliminate the contamination of the spatial mapping. We used a ZEISS CIRRUS system for data acquisition. The OCTA data from 40 eyes of 20 healthy participants were randomly split into 38 training and 2 testing data sets.
Results :
Figure 1 is a demonstration of the high-resolution reconstruction in wide-field OCTA. (a) is the original retinal OCTA image. (b) is the deep-learning-reconstructed high-resolution OCTA image. (c), (d), (e), and (f) are the zoom-in views inside the colored boxes in (a). Significant improvement in resolution could be observed in the entire FOV.
Conclusions :
The proposed method could effectively solve the trade-off between resolution and FOV. It will benefit the accurate quantification of vascular biomarkers over a large FOV for assisting diagnosis and treatment.
This is a 2020 ARVO Annual Meeting abstract.