En face analysis of the short posterior ciliary arteries crossing the sclera to the choroid using wobble-source wide-field optical coherence tomography (2023)

This study was conducted under the Declaration of Helsinki and approved by the University of Pittsburgh Institutional Review Board. Informed consent was obtained for each healthy volunteer who underwent swept well OCT imaging. Each subject underwent a best-corrected visual acuity assessment, full ophthalmological examination, and a wide-field 12 × 12 mm SS-OCT in the Plex Elite 9000 device (Carl Zeiss Meditec, Dublin, CA) centered on the fovea. Searched OCT source scans were exported as full 8-bit volumes. Each OCT volume consists of 1024 B-scans, each with a resolution of 1024 × 1536. We included healthy subjects. Any subject with diabetes, high blood pressure, or any systemic disease or treatment was excluded. Any subject with a history of maculopathy or retinopathy and severely myopic eyes was excluded.

photo analysis

Estimation of the inner (CIB) and outer (COB) boundaries of the choroid

Structures in the OCT volume, including the sclera, appear curved due to the bulbous shape of the eye. Therefore, to obtain parts of the sclera, it is necessary to restructure the OCT volumes to flatten the sclera. In other words, the scleral area below the choroidal-scleral interface (CSI) must be flattened. To that end, COB detection is a critical step in locating the sclera below the choroid. We adopted a modified version of our previously described method²³Segment COB. The previous method had two major limitations: (i) the detection of CIB and COB was based on two separate approaches and the structural similarity (SSIM) technique previously used to obtain the initial COB estimate was computationally intensive; and (ii) individual B-scan processing was performed and the inter-slice dependence within the B-scans of an OCT volume was not considered in the detection of CIB and COB, which may not result in seamless (spatially consistent) boundaries over B - scans away. With this in mind, we have made two new contributions to the modified methodology to address the above limitations. In particular, this includes (i) obtaining first estimates of both CIB and COB based on a single technique, i.e. H. using a two-step potentiation enhancement method inspired by the peculiarities of OCT imaging; and (ii) performing volumetric smoothing to correct misrecognitions in initial estimates of CIB and COB, taking advantage of the inherent neighborhood dependence between the B-scans of an OCT volume. The scheme of the proposed method for detecting CIB and COB is shown in Fig.3The corresponding implementation output of each step is shown in Fig.4based on a representative OCT-B scan (Fig.4b) an OCT volume (Fig.4A). The coordinate orientation of the XYZ axis of the OCT volume is also shown in Fig.3A. In particular, the proposed method includes (i) sequentially detecting the initial CIB estimate in each B-scan and stacking all estimates in 3D; (ii) obtaining the final estimate of the CIB by smoothing in a direction perpendicular to the B-scan alignment; (iii) sequentially recognizing the initial COB estimate in each B-scan and stacking all estimates in 3D; and (iv) obtaining the final COB estimate by applying step (ii) from CIB to COB. Below are step-by-step details of the method.

preprocessing

In OCT imaging, signal attenuation usually occurs in deeper regions, resulting in poor contrast between the vascular region and the stromal region in the choroidal layer. In addition, there is inherent speckle noise due to the coherence of light. To overcome the above limitations and improve the signal-to-noise ratio, we apply certain pre-processing steps. In particular, second-order median filtering with 6 × 6 tiles and adaptive histogram smoothing are applied to the B-scans to reduce the inherent speckle noise and improve the contrast of the image, respectively (Fig.4CD). In addition, the exponential gain used in the following steps compensates for the signal attenuation.

Common improvement method

The main techniques for obtaining first estimates of CIB and COB are exponentiation and nonlinear enhancement. We used these edits to binarize the OCT-B scan to quantify the choroidal stroma-luminal area²⁴. These operations are performed on the raw OCT intensity scale. Correspondingly for a B-scan image with grayscale intensity\(I,\)Raw intensity for each pixel\(\left( {x,y} \right)\)is obtained by

The exponentially improved image of\(I_{raw}\)is obtained by²⁵.

Wo\(N\)is the exponent and we chose empirically\(n = 10\). In general, the exponentiation gain compensates for the signal attenuation that occurs in deeper layers of the posterior segment during OCT imaging.

In addition, the non-linear enhanced image of\(I_{expenh}\)has been given by

The above operations used in estimating CIB differ slightly from the operations used in estimating COB. Specifically, for CIB detection they are applied to a vertically flipped B-scan to highlight retinal structures, while for COB detection they are applied to the original B-scan alignment to highlight choroidal structures. The details are described below.

Detection of the inner edge of the choroid (CIB).

CIB detection is performed to locate the choroidal region below the retina and facilitate COB detection. As mentioned before, we first get an initial CIB estimate for each B-scan and then smooth it over the B-scans to get a final, smoothed CIB estimate.

First estimate CIBWe used exponential and nonlinear enhancement operations (Eq.(1)–(3)) on a mirrored OCT image (Fig.4e) to strengthen the retinal substructures while reducing the intensities corresponding to the choroid and sclera (Fig.4F). In particular, the intensities of the retinal pigment epithelium (RPE) and other retinal layers become saturated at the bottom of the image due to their higher row indices. This is a reverse operation that has been described and previously reported for estimating choroidal stromal-luminal ratio²⁴. Thresholding is then performed using an average gray level intensity threshold of 128 to obtain a binarized image. This results in closely spaced large contiguous components of the retinal area. There may also be some minor spurious components. Accordingly, morphological operations, including the closed, open, and connected component algorithm, are performed to detect spurious components and to detect only large connected components belonging to the retinal layers (Fig.4g) is busy. Then to obtain the first CIB estimate (Fig.4Hi). Note that the above operations are reused in the following steps when detecting COB.

Flatten in the orthogonal directionIn the initial CIB estimates of some OCT volume B scans, we noticed some misdetections, especially when the retinal pigment epithelium (RPE) is deformed or depleted. With this in mind, we proposed to smooth the stack of all initial CIB volume estimates in the orthogonal direction (\(x{\text{-}}z\)level) to the level with the B-scans. In particular, we want to use the neighborhood interdependence between the boundary estimates to correct for the discrepancies. For each set of CIB estimates in the orthogonal direction, we applied robust locally estimated scatterplot smoothing (RLOESS).²⁶. Then we applied tensor voices^23,27(adapted from our previous work²³) in the original B-scan direction (\(y{\text{-}}z\)plane) to compensate for small distortions that make up the final CIB estimate (Fig.4J).

Detection of choroidal outer boundary (COB).

COB detection also includes B-scan processing in two stages, including scan-by-scan detection of initial COB estimates and correction of any noise detections in the volume by smoothing in the direction perpendicular to the plane containing the B-scans .

First COB detectionIn our previous work, we mainly used the index part of structural similarity (SSIM) through exponentiation and nonlinear enhancement operations (Eq.(1)–(3)). However, unlike CIB detection, they are applied in the original orientation without flipping the B-scan (Fig.4k,n). Then tensor tuning^23,27is used to generate a continuous smooth initial COB estimate (Fig.4O).

Flatten in the orthogonal directionSimilar to the initial CIB estimate, we found that the initial COB estimate deviated from the true limit in some B-scans. Therefore, we re-run RLOESS to resolve these outliers²⁶Flattening the stack of initial COB estimates in the direction perpendicular to the plane containing the OCT B scans (Fig.4P).

The detected CIB and COB values ​​for the entire considered OCT volume are shown in Fig.4Q.

Sclera-One-Face Image Extraction

As mentioned earlier, the structures in the OCT volume, including the sclera, assume a spherical shape due to the anatomical structure of the eye. Therefore, it is necessary to flatten the COB (choroid-sclera interface) to obtain scleral enface sections. To do this, after determining the COBs for the entire OCT volume, the pixel coordinates of the OCT volume are transformed so that the COBs of all B-scans lie on one plane. For the representative OCT volume, Fig.5shows the OCT volume with flattened COB, the OCT volume with detached retinal and choroidal layers, and finally the extracted sclera-and-face image.

Validation of the choroidal segmentation algorithm

The choroidal segmentation algorithm was validated separately for the choroidal inner boundary (CIB) and the choroidal outer boundary (COB) using 100 scans randomly selected from each of the five datasets (500 scans in total). The performance of automated detection of CIB and COB was validated by considering intraobserver variability as a reference. Manual demarcations were performed twice by an expert using ImageJ software. Boundaries created in the first round are masked in the second marking round. In particular, an initial piecewise linear boundary was first marked by using the Segmented Line tool to select sparse spots on the boundary. A smooth boundary was then obtained by splining the original boundary using the Fit Spline edit (Edit → Selection → Fit Spline).

Cube coefficient (DC) for choroidal segmentation validation

We used the proven Dice Coefficient (DC) metric to compare the accuracy of CIB and COB algorithmic segmentation with corresponding manual segmentations. It is defined below.

Permit\(x_{ij}\)In\(y_{ij}\)represents the two thickness measurements on the i-th\(\links( {i = 1, \ldots,N} \rechts)\)A-scans2 of the jth B-scan. In addition, the pixel indexes match\(x_{ij}\)In\(y_{ij}\)are represented by\(C_{xij}^{{}} \left( {\left| {C_{xij}^{{}} } \right| = x_{ij} } \right)\)In\(C_{yij}^{{}} \left( {\left| {C_{yij}^{{}} } \right| = y_{ij} } \right)\). Then the cube coefficient (DC) at the A scan location\(ij\)is determined by

A value of DC = 1 (100%) indicates that two measurements are equal, and DC = 0 indicates that two measurements are completely uncorrelated.

We considered the DC between the two manual segmentations as a measure and compared it with the DC between the proposed algorithmic segmentation and the reference manual segmentation obtained by averaging two manual segmentations.

localization of the arteries

The OCT volumes were then rebuilt with custom A scan positionszto level the chorio-scleral junction. A substack of 5 to 30 facial projections was manually selected to select the best stacks showing the arteries.

The observer manually marked the arteries that could be traced between the stacks and emptied into the choroid. Nine square squares (each 4 × 4 mm).²) were determined based on the facial image. The number of PCAs in each sector was determined based on their location. Measurements of the ships in this way were made twice by the two masked observers. Masked observers were masked both for their initial measurements and for other observers' measurements. Eyes with fewer than 10 PCAs identified due to low contrast were secondarily excluded from the analysis as this represents the minimum number of SPCAs identified in histological studies¹².

Statistical analysis

Correlation analysis (e.g., Pearson correlation) was used to determine choroidal vessel density^3,4,5and development of a feature learning technique²⁸. In our approach, interclass (cross-correlation) and intraclass (autocorrelation) correlations were determined to detect the SPCA entry site. Statistical analysis was performed using ExcelStats. Suppose X and Y are the pixel-based position information of an artery on a 2D image plane by considering the lower left corner as the origin of the image pixels in a Cartesian coordinate system. We then define Z as the positional information of the artery so determined by a human expert

Wo\(\Epsilon\)determines the robustness of artery detection. We then used the correlation between Z_Tin Z_t+1as a measure to determine the arterial positions on an image, where Z_Tin Z_t+1Satisfy the above equation at trial t with n images (we used n = 18).

Then the Pearson correlation is calculated in auto-correlation and cross-correlation. The qualification of the correlation is based on the empirical characterization of the Divaries correlation coefficient^29,30, which defines the following groups:

• below 0.20 → very weak correlation;

• 0.20 to 0.40 → weak correlation;

• 0.40 to 0.60 → moderate correlation;

• 0.60 to 0.80 → strong correlation;

• above 0.80 very strong correlation.

The distribution of arteries (percentages) was analyzed in 9 sectors (each square = 4 × 4 mm²) above the back pole.