Can an automated algorithm identify choriocapillaris in 2D-optical coherence tomography images?

Abstract

This editorial considers the problem of automatic identification of choriocapillaris in 2D-optical coherence tomography images. Firstly, the major challenges in automating the process are presented. Subsequently, the authors present an overview of current methods in the automatic segmentation of 2D-optical coherence tomography images. Lastly, the authors close with a discussion on a recent algorithm for the automated identification of choriocapillaris based on the identification of the retinal pigment epithelial layer and the region in its immediate neighborhood, and its capabilities.

Keywords::

• automation
• choriocapillaris
• OCT
• RPE
• segmentation

The choroid is the most vascular part of the eye, which supplies blood to the outer retinal structures. It consists of three layers – the choriocapillaris, medium-sized vessels (Sattler’s layer) and large-sized vessels (Haller’s layer). Among these layers, the choriocapillaris that lies just below the retinal pigment epithelial (RPE) layer is most important in the microcirculatory anatomy of retina.

The choriocapillaris, or any other choroidal structures, cannot be visualized through fundus examination. Previously available imaging modalities such as ultrasonography and indocyanine green angiography could demonstrate changes in the choroid in various retinal diseases [1–3]; however, due to their inability to evaluate the choroid quantitatively, combined with the low resolution of these devices, choroidal assessment was incomplete. Recent advances in spectral domain optical coherence tomography (SD-OCT) have evolved to image the choroid in vivo and is now able to provide structural details of each layer [4,5].

With light microscopic examination, Ramrattan et al. showed age-related change in choriocapillaris density and diameter, which was significantly larger in age-related macular degeneration (AMD) eyes compared with normal macula [6]. In addition to this finding, atrophy of choriocapillaris has been postulated to be an important factor in the pathogenesis of AMD [7,8]. Therefore, early in vivo changes in choriocapillaris assessment could be useful in the diagnosis and management of AMD.

Enhanced depth imaging is one such tool that is capable of providing quantitative assessment of the choroid including choroidal thickness, choroidal volume and individual choroidal vessel layer thickness measurements. The techniques include manual demarcation of the choroid; however, manual demarcation is always time-consuming and needs training for the observers [9–12]. Branchini et al. analyzed individual layers of the choroid using manual segmentation; however, they included choriocapillaris in the Sattler’s layer, as due to its very small size, manual choriocapillaris segmentation is very difficult to achieve. Automated segmentation of choroid and its layer is still not available in any of the commercially available devices.

The need for automated segmentation of choroidal images cannot be overemphasized but is very challenging. We hereby present some of the challenges involved in the automated segmentation of the choroid. Image segmentation algorithms typically assume reasonable structural demarcation (i.e., good edge information). In other words, edges are assumed to be too sharp and have an acceptable signal-to-noise ratio (SNR). Also, edges are assumed to not have many discontinuities.

In a typical SD-OCT image, the RPE layer is the brightest and all other information-conveying intensities are significantly darker. SNR of the image varies significantly from bright to dark regions. This is especially true for choroid–scleral boundary. The SNR also depends on the device and its operational settings. Also, edges are not typically strong and contain several discontinuities. Given the complex network of vessels, parameterizing their shapes cannot be easily achieved either. Therefore, a majority of the assumptions made by image segmentation algorithms are clearly violated in typical SD-OCT images, thereby making their segmentation a nontrivial task.

There are many reports, which include automated demarcation of choroid, that address some of the challenges described above. The approaches can be broadly classified into gradient-based techniques, statistical techniques and a combination of the two.

Kajić et al. proposed a two-stage statistical model to detect choroid boundaries in the 1060 nm OCT images in healthy and pathological eyes [13]. The first stage carries out initial boundary detection, while the second stage builds a statistical model of the expected shapes and their spatial relationships. The statistical modeling provides robustness to changes in SNR and shape properties. This approach gives good results (mean error of 13%) at the cost of requiring extensive training for the statistical model.

Alonso-Caneiro et al. reported an algorithm for the segmentation of the inner and outer choroidal boundaries in SD-OCT images [14]. The method used was graph search theory to segment the two boundaries of interest, applying two different strategies to calculate the graph weight maps. For inner boundary detection, the algorithm uses an edge filter and a directional weight map penalty, while for the outer boundary detection, algorithm is based on OCT choroidal image enhancement and dual brightness probability gradient information.

Tian et al. found the choroidal boundary by finding the shortest path of the graph formed by valley pixels using dynamic programming [15]. The average of Dice’s coefficient on 45 enhanced depth imaging OCT images was 90.5%.

Lu et al. proposed a technique to segment the inner boundary of the choroid using a two-stage fast active contour model [16]. Then, a real-time human-supervised automated segmentation on the outer boundary of the choroid was applied. The reported Dice similarity coefficient value on 30 images captured from patients diagnosed with diabetes was 92.7%.

Danesh et al. proposed a texture-based algorithm for fully automatic segmentation of choroidal images [17]. Dynamic programming is utilized to determine the location of the RPE. Bruch’s membrane was segmented by searching for the pixels with the biggest gradient value below the RPE. Choroid–sclera interface was segmented using wavelet-based features to construct a Gaussian mixture model. The model is then used in a graph cut for segmentation of the choroidal boundary. The main limitation of this method is the need for manual segmentation to construct the model.

Zhang et al. reported automated choroidal vasculature thickness and choriocapillaris-equivalent thickness of the macula using 3D SD-OCT [18]. Choroidal vessel segmentation included two main steps: vessel detection and vessel segmentation. In the vessel detection step, choroidal vessels were modeled as 3D tube-like objects in a resampled subvolume that yielded isometric (cubic) voxels using multiscale Hessian matrix analysis. The vesselness map of the choroidal vasculature then was calculated using the eigenvectors of the tensor matrix at each voxel position. For the second main step of vessel segmentation, voxel groups with relatively high vesselness values were selected by thresholding of the vesselness map using an experimentally determined threshold that was fixed for all analyzed images. Varying this threshold gave close-to-equivalent results over a large range of values. The resulting binary regions were used as seeds for a classic region-growing approach to segment the choroidal vasculature.

Zhang et al. considered choroidal thickness as the thickness of the choroidal vasculature, that is, upper surface of the vessel below the Bruch’s membrane and lower surface of the choroidal vessel above the sclera. Choriocapillaris thickness was measured between the Bruch’s membrane and upper surface of the choroidal vessel. However, the authors acknowledged the limitation of the study as inability to measure the outer wall of the choroidal vessel due to poor SD-OCT quality.

We reported an automated two-stage algorithm for the identification of the choriocapillaris region in 2D-OCT images [19]. In the first stage, the algorithm identifies the bright RPE cell layer to segment the image and identify the region of interest. In the second stage, the algorithm uses a gradient-based method to identify the upper and lower edges of the choriocapillaris region. The second stage is guided by the structure of the choroid and its relation to the RPE. The algorithm is able to identify the choriocapillaris well but relies on empirically determined thresholds for its successful operation.

Overall, the automated segmentation of choroid including choriocapillaris is under research. Quantification of the damage and more objective assessment would improve the diagnostic abilities as well as follow-up in various chorioretinal disorders. Further research in automated assessment of choroid is warranted to improve its clinical application for choroidal vascular disease diagnosis and management.

Financial & competing interests disclosure

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.