Before segmentation, images were median filtered by 3 pixels in the transverse direction and down sampled to consist of 500 transverse pixels. Because of the high axial scan density, this step did not result in a loss of transverse resolution. All boundaries except the ELM had to occur at 0 crossings of the second derivative of the linear image. Therefore, the following criterion was used:
\[s_{\mathrm{zc}}{=}\mathrm{zero\ crossings}(g{^{\prime\prime}}{\ast}A),\]
where
g is a Gaussian kernel, where the optimal width of the kernel was determined empirically for each boundary and was the same for all measured eyes.
A is the median filtered axial scan. The overall edge strength criterion for each boundary except the ELM could be expressed as
\[s_{\mathrm{zc}}{=}\mathrm{zero\ crossings}(g{^{\prime\prime}}{\ast}A)s{=}s_{\mathrm{zc}}{\times}({\pm}g{^\prime}{\ast}A){\times}(h_{1}{\ast}A)^{n_{1}}{\times}(h_{2}{\ast}A)^{n_{2}}.\]
For each axial scan, the maximum value of
s determined the initial estimate for the boundary. In the expression, * denotes convolution along the axial direction, × denotes multiplication, and
s is the product of four terms, which represent the four criteria that must be satisfied for the boundary of interest. The first-term
s zc selects 0 crossings of
g′′ *
A. The second term ±
g′ *
A weights regions where the derivative along the axial direction is large in magnitude and has the desired sign. The positive sign is chosen for positive-going boundaries (VRI, IS/OS inner boundary, and COSTs boundary), and the negative sign is chosen for negative-going boundaries (IS/OS outer boundary and BM outer boundary). The kernels
h 1 and
h 2 are causal and anticausal functions used to select regions proximal and distal the boundary, respectively. Depending on the signs of
n 1 and
n 2, the last two terms weight boundaries near regions of either high (positive sign) or low reflectivity (negative sign). The magnitudes of
n 1 and
n 2 determine the relative weighting of the different terms. For instance, the VRI boundary must be distal to a region of low reflectivity corresponding to the vitreous; therefore, for this case
n 1 = −4 and
h 1 was a causal rectangular kernel. The VRI boundary must be proximal to a region of high reflectivity corresponding mostly to the nerve fiber layer (NFL), therefore for this case
n 2 = 1 and
h 2 was an anticausal rectangular kernel. The shapes of the kernels
h 1 and
h 2 and the values of
n 1 and
n 2 were determined empirically for each boundary and were the same for all measured eyes. In order to detect the ELM, the median-filtered image was used directly to search for a thin reflective line between 10 and 35 μm proximal to the IS/OS junction.