2.1 PS-OCT Instrument
A spectral domain PS-OCT setup was used for imaging, a scheme of which is shown in . A superluminescent diode (Superlum Diodes, Inc.) centered at 839 nm with a FWHM bandwidth of 58 nm was used as a light source. The light was inserted into a Michelson interferometer in a vertically polarized state using a bulk optics polarizer. After passing a polarizing beamsplitter (PBS), the light was split into a sample and a reference beam by means of a nonpolarizing beamsplitter (BS). In the sample arm, a quarter wave plate (QWP) was used in order to provide circularly polarized light onto the eye. Further sample arm optics included an X
galvanometer scanner and a relay of two identical achromat lenses. A probe power of 700 μ
W on the cornea was used, which is well below the limits drawn by American National Standards Institute (ANSI) and International Electrotechnical Commission (IEC).32,33
Half of the light beam returning from the sample arm was again directed into the source arm. There, its horizontally polarized portion was detected by an avalanche photodiode (APD), in order to be able to record scanning laser ophthalmoscope (SLO) images using the same sample arm optics.
Fig. 1 Combined PS-OCT and SLO instrument for clinical imaging. Components: SLD, superluminescent diode; PC, polarization controller; FC, fiber coupler; POL, polarizer; PBS, polarizing beamsplitter; BS, nonpolarizing beamsplitter; QWP, quarter-wave-plate; GS, (more ...)
In the reference arm, the beam passed a variable neutral density filter, glass prisms for dispersion compensation, and was reflected by a mirror mounted on a motorized translation stage. Light returning from sample and reference arm, respectively, interfered at the nonpolarizing BS. Note that the QWP present in the reference arm of the original PS interferometer scheme devised by Hee et al.19
was omitted in our design in order to avoid saturation of the APD by light returning from the reference mirror. This altered layout required an additional half-wave plate (HWP) oriented at 22.5 deg in the detection arm, which rotated the beam’s polarization state by 45 deg in front of the PBS in the detection arm, such that equal reference arm power was provided in both orthogonal polarization channels. In a software compensation step, this rotation was compensated for by multiplying the OCT signals from the two channels with the Jones matrix of a HWP oriented at −45 deg. Now, the previously published method for computing PS-OCT signals could be applied.34
From the interferometer exit, the signals of the two orthogonal polarization channels were guided to two identical spectrometer units employing diffraction gratings with 1200 lines/mm and achromat lenses (f = 200 mm). The spectral interference signals were read out simultaneously at two line-scan cameras (e2v AViiVA M2 CL 2014) operating at 20 kHz. Each readout of the two cameras yielded the spectral data necessary to compute not only reflectivity but also phase retardation, axis orientation, and Stokes vectors for one depth profile.
Two operational modes were available for imaging, which were timed and controlled using a field-programmable gate array (FPGA). For alignment of the eye under investigation, SLO images and OCT B-scan images were recorded and displayed in an alternate fashion to provide an overview in three dimentions. SLO and OCT images were recorded in 165 and 20 ms, respectively. In the real-time display, each SLO image was followed by four OCT images. Once the eye was aligned properly, 3-D data sets covering a retinal cube of 6.2×6.7×3.3 mm3 (x×y×z) were recorded. One out of three sampling patterns (64×1024, 128×512, 256×256) could be selected. For this study, only data sets sampled 128×512 (that is, 128 B-scans each consisting of 512 A-lines) were used.
The system was used for imaging in the eyes of patients suffering from various ocular diseases. Compliant patients diagnosed with drusen and advanced dry AMD (geographic atrophy), respectively, were selected and imaged for this study. The study protocol was approved by the ethics committee of the Medical University of Vienna and followed the tenets of the declaration of Helsinki.
2.2 Degree of Polarization Uniformity
Segmentation of the RPE was performed based on its depolarizing (i.e., polarization scrambling), character as reported in Ref. 26
. For each B-scan, the signals recorded at the two exits of the PS-OCT instrument were processed using regular SD-OCT processing, including subtraction of an averaged spectrum, rescaling, numerical dispersion compensation, and Fourier transform.35
From the resulting signals, I1,2
) = A1,2
)], the Stokes vector S
= (I Q U V
was computed for each image pixel. In order to gate out noisy pixels, for further processing, only pixels with intensities above a depth-dependent threshold Ithr
) = Ithr
= 0 μ
m were considered. Ithr
= 0 μ
m) was set to a value of ~11 dB above a level equal to two standard deviations of the average noise amplitude in each A-scan.36
Each of the Stokes vector elements Q, U
, and V
was normalized by I
and then averaged in a sliding window sized 8(x
) pixels (~100×16 μ
). Thereof, the degree of polarization uniformity (DOPU) was calculated as
DOPU has a form similar to the degree of polarization known from conventional polarization optics. Dependent on the uniformity (or nonuniformity) of the polarization states of pixels inside of the respective evaluation window, DOPU values will exhibit values in the range from 0 to 1. DOPU values close to 1 will be found in polarization preserving and in birefringent structures whereas polarization scrambling (depolarization) will manifest in lower DOPU values.
Now, from each DOPU B-scan image, depolarizing structures can be derived by segmenting pixels with low DOPU values, that is, pixels with DOPU values below a threshold DOPUthr. Segmentation results may vary, depending on the choice of DOPUthr, and the optimal DOPUthr value can be different, depending on the respective eye and on the image quality. For the results and figures in this paper, a fixed value of 0.8 was chosen for DOPUthr.
shows PS-OCT B-scan images of a human retina in the fovea region of a healthy volunteer’s left eye. In , the reflectivity B-scan image is displayed where healthy layers can be distinguished based on the intensity of backscattered light. The respective DOPU image is shown in . While most layers—those appearing in reddish color—do not considerably alter the sample beam’s polarization state, polarization-scrambling manifests in lower DOPU values displayed in green and blue. Note that, due to the computation of DOPU by averaging in a sliding evaluation window, the spatial resolution is somewhat reduced in this image. By considering only pixels with low DOPU values, depolarizing structures can be segmented as shown in , where low-DOPU-valued pixels are overlaid in red. In , the position of the internal limiting membrane (ILM), which was found by intensity thresholding, is overlaid on the reflectivity image in blue. Furthermore, the position of the pixel with the lowest DOPU value in each A-line (i.e., the pixel exhibiting the highest depolarization), which in the healthy retina is located in the RPE, is displayed in red. A fundus image computed from the OCT data set by axial summation of the intensity for each A-line is shown in , and in , a retinal thickness map derived by calculating the axial distance between ILM and RPE position can be seen. An average refractive index of 1.38 was assumed in the retina.
Fig. 2 PS-OCT images. (a) Reflectivity B-scan image. (b) DOPU B-scan image [color scale: DOPU = 0 (dark blue) to DOPU = 1 (red)]. Pixels with intensities below a certain threshold are displayed in gray. (c) Overlay of depolarizing pixels on reflectivity image (more ...)
2.3 Segmentation of Drusen
Drusen were segmented by first locating the actual position of the RPE by its DOPU value. This line, found by tissue-specific contrast, acts as a “backbone” for the following calculation of the position where the RPE should be in a healthy eye. Finally, the drusen are segmented and quantified by calculating the difference between the actual position and the normal “should-be” position.
For each B-scan, the starting point of the calculation is the axial position ZRPE(i) of the minimum DOPU value in each A-line, that is, the red line in . The index i indicates the index of the A-line in the respective B-scan and ZRPE(i) is measured as the distance from the zero delay, which is located at the upper border of all shown B-scan images, and ZRPE(i) increases in the posterior direction. First, missing values of ZRPE(i) are interpolated in A-scans, where no pixel is below the threshold DOPUthr. Then, outliers of ZRPE(i) are cut and interpolated. Outliers are defined as values ZRPE(i), which differ by >33 μm (i.e., 10 pixels) either from the mean of ZRPE(i − 1) and ZRPE(i − 1) or from the mean of ZRPE(i + 1) and ZRPE(i + 2), or from both. If ZRPE(i) is an outlier, then ZRPE(i) is replaced by the average of its four neighbouring values ZRPE(i + m), where m = −2, −1, +1, +2. The subsequent function is smoothed using a Savitzky–Golay filter (polynomial order: 3, filter length: 6).
The result of these first steps is a function
without gaps or outliers. The following part of the computation is an iterative approximation of the normal RPE position. Every iteration of the loop comprises two steps: In the first step, the output Bj−1
of the previous loop iteration is used as an input. For the first iteration, the input A0
) = maxi
is used. Now,
is the axial extension of one pixel (~3.2 μ
m, and optical distance) j
[1,175] is the respective iteration number of the loop. The offset 10Δzpix
was chosen to compensate for flattening effect of the subsequent smoothing step.
The new function Aj
is used in the following second step to compute the output Bj
where a smoothing Savitzky–Golay (SG) filter is represented with a window length of 200 using a third-order polynomial. As the loop is iterated 175 times, Bj
converges toward a smooth function resembling the normal RPE position. As a final point, B175
yields the final function ZNRPE
. Here, the off-set 10Δzpix
accommodates for an axial shift of the smoothed function B175
Drusen can now be found by computing the difference between actual RPE position ZRPE and normal RPE position ZNRPE for each A-line. In the case of very large drusen or drusenoid pigment epithelium detachments in retinas appearing heavily bent in the OCT B-scan image, ZNRPE may be displaced in anterior direction (i.e., between the actual position of Bruch’s membrane and the RPE). This will not affect the calculated drusen area, but the volume may be underestimated. We observed this effect in confluent drusen with a diameter of ~3 mm. For the results presented in this study, which included patients with drusen classified as AREDS score 2 and 3, this effect was not apparent.
2.4 Segmentation of Geographic Atrophy
Using PS-OCT, atrophic zones in dry AMD can be detected by detecting holes in the depolarizing RPE layer.
In a straightforward manner, one could simply sum the number of depolarizing pixels [i.e., the red pixels in ] for each A-line of a 3-D PS-OCT data set. The resulting thickness map of the depolarizing layer now reveals the existence and absence of polarization scrambling tissue in an en face view.
However, depending on the subject’s pigmentation, depolarization may also be apparent in the choroid. Signals from this layer usually are rather weak when the probe beam anteriorly has to penetrate the highly scattering and absorbing RPE. Contrarily, in eyes with geographic atrophy, the missing RPE usually results in higher penetration depths (sometimes down to the sclera) and in stronger OCT signals from the choroid. Dependent on the subject, choroidal tissue exhibits a more or less polarization scrambling character. In some cases, depolarizing spots with DOPU values as low as those in RPE tissue may be noticeable. If, in such eyes, the number of all depolarizing pixels is integrated over depth in order to compute a thickness map of the RPE, then polarization scrambling pixels in the choroid may add to those in the RPE. This can result in mimicking of healthy structures and masking of RPE defects, if the RPE is absent or thinned as in the case of geographic atrophy.
To prevent from integrating choroidal pixels into computing of the thickness map of the depolarizing layer, only depolarizing pixels located in a shallow band close to the photoreceptor layer and Bruch’s membrane should be contributing.
For this purpose, the ILM position is used as a starting point. The initial ILM position is smoothed similar to the RPE position in the previous section by using a third-order polynomial SG filter with a window length of 200, yielding the function
. Now, starting from positions
, hyperintense structures are detected by seeking the first pixel with an intensity greater than the threshold computed as the average of the maximum pixel values of each A-line minus 7 dB. Typically, the positions in the photoreceptor/RPE complex or Bruch’s membrane are detected. Subsequently, discontinuities are interpolated and, similar to the iterative approximation of the normal RPE position in drusen described above, a smooth function is computed to closely fit to its posterior border. This resulting smooth function is used as the spine of the evaluation band, which extends 15 Δzpix
in the anterior and posterior directions. This width ensures that depolarizing spots in the choroid are excluded from the band, yet allowing for small irregularities of the axial RPE position such as small drusen because they are being found in dry AMD.
By summing the number of depolarizing pixels within the evaluation band along each A-line, an RPE thickness map is generated. In order to assess the size of atrophic zones, the thickness map is binarized and smoothed by removing isolated pixels using an erosion filter. In a last step, the algorithm automatically detects patches of connected pixels, which can now be converted into atrophic areas by scaling them with the known pixel area.