Purchase this article with an account.
Yonggang Shi, Jin Gahm, Amir H Kashani; Curvelet-based Vessel Enhancement for 3D OCT Angiography. Invest. Ophthalmol. Vis. Sci. 2017;58(8):648.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Recent studies have successfully used OCT Angiography (OCTA) to extract 2D metrics for the quantitative assessment of capillary morphology and density. While highly valuable, these 2D OCTA metrics were derived from the en face projection of the 3D OCTA data, and therefore inevitably obscure the geometric and topological information in the original 3D vasculature networks. To robustly compute truly 3D metrics for the automated analysis of OCTA data in large scale eye studies, it is important to first develop robust 3D vessel enhancement methods from OCTA data.
Conventional filtering methods tend to obscure small vessels when they try to suppress the noise in 3D OCTA data. In this work, we propose the use of a model-based filtering approach based on the curvelet transform (Candes E et al., 2010). The key idea is that the curvelet basis functions provide a multiscale representation of edges and singularities along curve-like structures in 3D images, which matches very well with the geometry of retinal vasculature in OCTA images. Before we process the OCTA image, the IOWA Reference Algorithm (Abramoff MD et al., 2010) is first applied to segment the retinal layers. For each 3D OCTA volume, we then apply the 3D wrapping-based curvelet transform at 5 scales with 8 orientations. Denoising is then achieved by adaptively thresholding the curvelet coefficients at multiple scales in proportion to the noise level, which is automatically computed from the outer nuclear layer (ONL).
To demonstrate our curvelet-based method for 3D OCTA analysis, we show the results from a normal control and a patient with diabetic retinopathy. The input data is the raw 3D image volume from the OCT scanner. Using the denoised 3D OCTA image, we generated en face Maximum Intensity Projection (MIP) images and color-coded the images with the depth of the projected vasculature. To achieve invariance with respect to retinal anatomy, we calculate the depth as the distance to the boundary of ONL. From the results shown in Figure 1, we can clearly see that our method is able to reconstruct the fine-grained retinal vasculature while removing noise in the OCTA data.
With our novel methods based on curvelet transforms, we can achieve high quality 3D donoising while preserving detailed retinal vasculature. This provides a foundation for further development of 3D OCTA analysis algorithms.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only