Visible light optical coherence tomography angiography (vis-OCTA) facilitates local microvascular oximetry in the human retina

: We report herein the ﬁrst visible light optical coherence tomography angiography (vis-OCTA) for human retinal imaging. Compared to the existing vis-OCT systems, we devised a spectrometer with a narrower bandwidth to increase the spectral power density for OCTA imaging, while retaining the major spectral contrast in the blood. We achieved a 100 kHz A-line rate, the fastest acquisition speed reported so far for human retinal vis-OCT. We rigorously optimized the imaging protocol such that a single acquisition took < 6 seconds with a ﬁeld of view (FOV) of 3 × 7.8 mm 2 . The angiography enables accurate localization of microvasculature down to the capillary level and thus enables oximetry at vessels < 100 µm in diameter. We demonstrated microvascular hemoglobin oxygen saturation (sO 2 ) at the feeding and draining vessels at the perifoveal region. The longitudinal repeatability was assessed by < 5% coeﬃcient of variation (CV). The unique capabilities of our vis-OCTA system may allow studies on the role of microvascular oxygen in various retinal pathologies.

The first demonstration of vis-OCT dated back to 2002 [23], but the first in vivo vis-OCT retinal imaging was not reported until 2011 [24]. On the other hand, the development of vis-OCT oximetry has made significant progress during the past 10 years. The feasibility of using visible light for oximetry was first demonstrated in vitro in 2010 [25,26], and in vivo in 2011 [6]. Subsequently, the first demonstration of vis-OCT retinal oximetry was reported in 2013 in rat retinas, showing the sO 2 calculations in the major arterioles and venules immediately exiting and entering the optic nerve head (ONH) [4]. When combined with the Doppler OCT measurements of blood flow, the oxygen metabolic rate from the inner retinal circulation could be quantified to access the global inner retinal functions and study the dynamic interaction between the retinal and choroidal circulations [3,15,27,28]. Later, human retinal imaging by vis-OCT was demonstrated in 2015 [29] and human retinal oximetry by vis-OCT was reported in 2017 [13,17] .
So far, most reports on vis-OCT human retinal oximetry describe vessels in close proximity with the ONH with vessel diameters above 100 µm [13,17] . The arteriovenous difference in those major vessels provides a global assessment of inner retinal metabolic activity but is less informative for localized metabolic alterations. For example, primary open-angle glaucoma (POAG) frequently manifests with visual field defects in the early stage only at the peripheral retina, gradually progressing towards the central visual field while diabetic retinopathy and macular degeneration often impact the foveal region initially. Therefore, measurement of localized inner retina metabolic change with focal oximetry in smaller vessels, even down to the capillary level, is highly desirable for better disease characterization.
Therefore, our overall motivation is to perform human retinal oximetry in retinal vessels less than100 µm in diameter by vis-OCT angiography (vis-OCTA). The paper consists of two major components. First, we implemented vis-OCTA in human retina to aid 3D localization of small vessels and capillaries. Given the sensitivity of our system, we optimized the scanning protocol to achieve a 3×7.8 mm 2 field of view (FOV). Second, we performed microvascular oximetry at the macular region, focusing on individual arterioles and venules immediately above the capillary network and then the capillary network itself after spatial averaging. The measured oximetry values were validated based on the known alternating pattern of feeding arterioles and draining venules in the macula. The sO 2 from the capillary network was found to be intermediate between measured values of the arterioles and venules.

System
The vis-OCT system was modified from our previous visible and near infrared OCT setup, with only the visible light channel operating ( Fig. 1(a)) [7]. Two edge filters selected the illumination wavelength range from 545 to 580 nm ( Fig. 1(b), 1(c)). We intentionally used a narrower bandwidth than our previous design to increase the power spectral density for subsequent vis-OCTA processing. Within the bandwidth, the large spectral contrast between oxy-and deoxy-blood can still be captured ( Fig. 1(b)). The sample arm consisted of a fiber collimator (f = 7.5 mm), a pair of galvanometer mirrors (GVS002, Thorlabs, USA), and a 2:1 telescope relay system. We also installed an electronically tunable lens (EL-3-10, Optotune AG, Switzerland) for a fast-focusing adjustment. We used a custom-made spectrometer to record the spectral interferograms. The spectrometer consisted of a fiber collimator (f = 60 mm), a 2400 lines/mm transmission diffraction grating (Wasatch Photonics, USA), a focusing lens (f = 150 mm), and a high-speed line scan camera (E2 V Octoplus, Teledyne, France). The laser power on the pupil was less than 0.25 mW, and the A-line speed was 100 kHz with an exposure time of 9.7 µs for each A-line. We note that Chong et. al., and Zhang et. al. used 0.15 and 0.08 mW power at the cornea, respectively, but for vis-OCT structural human retinal imaging [17,30]. We characterized the roll-off performance of the system using a mirror reflectance as the sample, and an ND filter (OD=2) to attenuate the sample signal ( Fig. 1(d)). The sensitivity was estimated to be ∼82 dB and the roll-off speed was ∼6 dB/mm in air. The axial resolution was measured to be ∼5 µm close to the zero-delay line.
All the experimental procedures were approved by the Boston Medical Center Institutional Review Board and adhered to the tenets of the Declaration of Helsinki. Informed consent was obtained prior to imaging.

Vis-OCTA imaging and processing
The scanning protocol is conventional and modified from our previous paper [31]. The fast scanning galvanometer mirror was controlled by an 80% duty cycle sawtooth wave. The slow scanning was controlled by a ramping wave. At each slow scanning location, repetitive frames were taken to harness the motion contrast of blood flow for vis-OCTA. There are three essential parameters in the imaging protocol: numbers of repetitive frames (N rep ), the time interval between two adjacent frames (∆T) for the motion contrast, and the A-line scanning density on the retina (Σ Aline , µm −1 ). These three parameters altogether determine the vis-OCTA image quality and the FOV, and their optimization is discussed in the Results section.
The method to generate wavelength-dependent four-dimensional (x, y, z, λ) vis-OCTA and vis-OCT has been described before [27,32]. We combined the split-spectrum OCTA and the complex-signal-based optical microangiography to enhance the motion contrast [33,34]. The split-spectrum method also provides the spectral information for sO 2 calculation.
The original raw spectrograms were pre-processed by normalizing the source spectrum, a DC removal, k-space resampling, and a digital dispersion compensation for improving the image sharpness [35]. A Gaussian spectral window in the wavenumber k domain (FWHM k = 0.11 µm −1 ) was then used to sweep the whole interferogram in 11 steps to generate the wavelength-dependent complex OCT signal after a series of Fourier transforms. The differential contrast between two adjacent frames at one slow scanning location was taken based on the complex OCT signal, and then the absolute value was taken as the OCTA signal. When N rep is larger than 2, all the OCTA signals from two adjacent B-scans were averaged to assemble an OCTA frame.
where C(x, z, λ i ) is the complex B-scan image at each spectral window λ i . After vis-OCTA were generated for each Gaussian spectral window, we averaged them to compile one 3D vis-OCTA image.
For wavelength-dependent vis-OCT structural images, we simply average the absolute values of N rep frames at each B-scan location, instead of taking differential contrast as in OCTA:

Data processing for oximetry in small vessels
We used the following workflow to extract the microvascular spectral signal and calculate the sO 2 ( Fig. 2(a)). The first step is to generate wavelength-dependent vis-OCT and vis-OCTA four dimensional (x, y, z, λ i ) datasets ( Fig. 2(b)), as described above. The second step is to locate the blood vessels through en face and axial segmentations. The third step is spectral normalization and extraction to estimate sO 2 .

Vessel segmentation
For individual arterioles and venules, the en face maximum intensity projection (MIP) of vis-OCT was used to locate the lateral region of interest (ROI) for the vessels. The vessel ROI was manually segmented and stored as the vessel mask ( Fig. 2(c)). We detected the retinal surface by setting an intensity threshold from the structural vis-OCT B-scans. The same surface was used in both vis-OCT and vis-OCTA datasets. We then flattened the retina surface and located the vessel bottom by the lower boundary of the vis-OCTA signal within each vessel segment (Fig. 2(c)). For each vessel segment, a constant depth selection was used for the spectral extraction assuming a consistent vessel size. The capillary flow signal from the en face MIP of vis-OCTA was used to isolate the capillary network. The larger vessels were marked out as shown in Fig. 2(e). Then we binarized the image to serve as the capillary ROI. To locate the depth of the capillaries, the axial position of the maximum value along each vis-OCTA A-line within the capillary ROI was recorded.

Spectral extraction and normalization
After 3D segmentation, a spectrogram in terms of depth was generated. We emphasize that the spectrogram was extracted from vis-OCT structural data instead of angiographic data. Vis-OCTA is only used to locate the vessels in 3D. The processing for arterioles and venules, and capillaries are described separately as below.
Arterioles and venules For individual arterioles and venules, the vessel surface was flattened and all vis-OCT A-lines within the vessel lateral ROI were averaged to generate a spectrogram in z and λ ( Fig. 2(d)). Then the spectrum was obtained by averaging signals from a depth range of 10 µm centered at the vessel bottom. To remove systemic bias due to propagation through the eye, we further performed a spectral normalization by the spectrum from the non-vascular nerve fiber layer tissues [17,36].
Capillaries For capillaries, we flattened all the capillaries and averaged the signal within the capillary lateral ROI to generate the spectrogram in z and λ as shown in Fig. 2(f). The signals from a depth range of 10 µm centered on the capillaries were averaged to extract the spectrum, which was further normalized by the spectrum averaged from the inner plexiform layer and the inner nuclear layer excluding the capillaries.

Spectral fitting to estimate sO 2
After the spectra were extracted and normalized, we used a previously published method to calculate sO 2 by matching the experimental spectra with the following model [4,27]: in which I 0 (λ) is the spectrum of the light source; R 0 is the reflectance of the reference arm and assumed to be a constant. r(λ) (dimensionless) is the reflectance at the vessel wall, modelled by a power law r(λ) = Aλ −α , with A being a dimensionless constant and α modelling the decaying scattering spectrum from the vessel wall. The optical attenuation coefficient µ (mm −1 ) combines the absorption (µ a ) and scattering coefficients (µ s ) of the whole blood, which is both wavelengthand sO 2 -dependent: where W is a scaling factor for the scattering coefficient at W = 0.2 [4,25]. Hb and HbO 2 denote the contribution from the deoxygenated and oxygenated blood, respectively. z denotes the light-penetration length through vessels. When taking a logarithmic operation, Equ. 3 turns to a linear equation to µ, log I(sO2|λ, z) Then a least square fitting was performed to match experimental spectrum I(λ) and model-based spectra to estimate sO 2 by optimizing a vector x, so that sO 2 is calculated by sO 2 =x 3 /(x 3 +x 4 ); and x 1 = ½ log(A), x 2 = ½ α, x 3 = sO 2 ×z, x 4 = (1-sO 2 )×z according to Eq. (5). The common variable z is removed by taking the ratio. We used the nonnegative linear least-squares fitting function lsqnonneg in Matlab. The values of sO 2 and α are reported in the result section. We should note that because of the additional spectral normalization by the retinal tissue, x 1 is found to be zero for all our spectral fittings, and the physical definition for α in Eq. (5) may not hold.

Imaging protocol optimization
We first evaluated the effect of N rep on vis-OCTA's signal-to-noise ratio (SNR), defined by SNR = I v SD nv (7) where I v is the vis-OCTA intensity on blood vessels, and SD nv is the image intensity standard deviation within the background non-vascular areas [37]. We collected a dataset by N rep = 12, and processed vis-OCTA images from the first N rep frames for N rep = 2, 4, 6, 8, 12. For example, the case of N rep = 2 only processed the first two frames and so on. This approach ensures that the other experimental conditions are virtually identical. We can see the image quality improvement when averaging more vis-OCTA frames ( Fig. 3(a)-3(e)). The image SNR increases more rapidly at the beginning and slows down when N rep > 6 ( Fig. 3(f)). While the image quality improves with increasing N rep , the acquisition time proportionally increases as well. Therefore, we chose N rep = 3 for the final imaging protocol to balance the image SNR and the total scanning time. The image density in Fig. 3 is 320 by 128 pixels. We next evaluated the effect of ∆T on the image quality. The A-line scanning density was kept constant at 0.13 µm −1 , while increasing ∆T from 1 ms to 7 ms by increasing the number of A-lines per frames from 80 to 560. As a result, the length in the fast scanning direction increased as well (Fig. 4(a)-4(d)) from 0.6 to 4.3 mm, respectively. The same FOV was imaged in all cases in Fig. 4(a), so that the SNR calculation could be repeated at the same imaging location. We noticed that the image contrast changes non-monotonically, increasing until ∆T = 5 ms and decreasing at ∆T = 7 ms. The initial SNR increase may be due to the larger decorrelation between two repeated frames when ∆T increases. The contrast then deteriorates when ∆T becomes longer than 5 ms as clear motion artifacts start to appear (e.g. vertical stripes in Fig. 4(c), 4(d)). To balance the FOV and the image quality, we chose ∆T = 5 ms for our imaging protocol. At 100 kHz A-line rate and 80% duty cycle of the fast scanning, ∆T = 5 ms determines the number of A-line per B-scan in our protocol to be 400, resulting in a ∼3 mm scanning range in fast axis.  Lastly, we evaluated the effect of A-line scanning density on the vis-OCTA SNR. We kept N rep and ∆T constant at N rep = 3 and ∆T = 4 ms with 320 A-line per B-scan. The length of the scanning was controlled by the amplitude of the sawtooth wave fed to the fast scanning galvanometer mirror. Based on the scanning angle at the pupil entrance and the approximated eyeball diameter of ∼25 mm, we varied the scanning length from approximately 0.375 to 3.375 mm in five acquisitions, corresponding to the A-line scanning density Σ A−line = 0.85, 0.29, 0.17, 0.12 and 0.09 µm −1 . As Fig. 5 shows, the image SNR decreases with the decreased A-line scanning density. This effect is likely due to the fringe washout effect as each exposure integrate signals from the scanning length per pixel and becomes more prominent as the scanning length increases. To balance the FOV, image quality and acquisition time, we chose Σ A−line = 0.13 µm −1 , ∆T = 5 ms, and 400 A-lines per frame for the final imaging protocol, which yielded an imaging length on the fast scanning direction of ∼ 3 mm.

Human retina vis-OCTA
After optimizing the protocol, we performed vis-OCTA on a human volunteer as shown in Fig. 6.The FOV at one single acquisition took 400 × 400 pixels to cover ∼3×7.8 mm in the retina. The N rep = 3 and the total time for one acquisition was 6 s. Figure 6(a) shows a wide field vis-OCTA mosaic including 11 acquisitions. Detailed microvasculature down to individual capillary-level can be visualized in two magnified images (Fig. 6(b),(c)). After the split-spectra processing, the depth resolution was relaxed to ∼27 µm, sufficient to isolate the superficial, intermediate, and deep capillary plexuses at the macula (see appendix, Fig. 8). The capillary contrast is higher within the nerve fiber arcade, which is consistent with the high metabolic demand in the retinal nerve fiber layer and the ganglion cells.

Microvascular sO 2 in human retina in vivo
We performed vis-OCT microvascular oximetry at the foveal region as demonstrated in Fig. 7. We color-coded the vis-OCTA images with sO 2 values for small vessel segments and the whole capillary bed (Fig. 7(a)). The alternating pattern of feeding arterioles and draining venules is apparent around fovea as demonstrated by the sO 2 estimations. As expected, the averaged sO 2 value over the capillary bed is intermediate between the values of the arterioles and venules. We plotted a representative spectrogram from an arterial segment, a venule segment, and the underlying capillary networks (Fig. 7(b)-7(d)). Figures 7(b) and 7(c) both have higher intensity at the vessel surface and this intensity attenuates when penetrating through the vessels. The depth at the vessel bottom provides sufficient signal contrast to extract reliable spectra. Figure 7(d) exhibits a different pattern with the strongest signal appearing within the capillaries as expected based on the stronger capillary contrast in structural vis-OCT images [30,38] .
In Fig. 7(e)-7(g), the spectra from all the arterioles, venules, and capillary beds displayed in Fig. 7(a) are plotted. The spectral contrast of the blood resembles that seen with the oxygenated blood plotted in Fig. 1(b), showing a peak around 560 nm. The value of sO 2 estimation is then plotted in Fig. 7 h as the 1 st subject. We further performed the same method on two additional value over the capillary bed is intermediate between the values of the arterioles and venules. We plotted a representative spectrogram from an arterial segment, a venule segment, and the underlying capillary networks (Fig. 6(b)-6(d)). Figures 6(b) and 6(c) both have higher intensity at the vessel surface and this intensity attenuates when penetrating through the vessels. The depth at the vessel bottom provides sufficient signal contrast to extract reliable spectra. Figure 6(d) exhibits a different pattern with the strongest signal appearing within the capillaries as expected based on the stronger capillary contrast in structural vis-OCT images [30,38] .
In Fig. 6(e)-6(g), the spectra from all the arterioles, venules, and capillary beds displayed in Fig. 6(a) are plotted. The spectral contrast of the blood resembles that seen with the oxygenated blood plotted in Fig. 1(b), showing a peak around 560 nm. The value of sO 2 estimation is then plotted in Fig. 6 h as the 1 st subject. We further performed the same method on two additional  healthy eyes, and their results are also summarized in the same graph, and Table 1. Overall, the mean values of arterial, venous, and capillary sO 2 are quite consistent among three different subjects. We also tested the repeatability of the microvascular sO 2 by vis-OCTA ( Fig. 7(i),). A total of 5 measurements were taken from the same eye on three different dates (see appendix, Table 2). The first three repetitions were taken on the same day, and the other two were taken on different days spanning over 5 weeks. The coefficients of variation (CV) for all measured vessels are < 3% within the same day and < 5% over all the repetitions.

Discussion
In this paper, we describe the first vis-OCTA imaging of human retina using a redesigned spectrometer and optimized imaging protocol allowing for enhanced vascular signal and localization of the microvasculature. Using our methods, we are capable of measuring sO 2 in vessels with diameter smaller than 100 µm, all the way down to the capillary networks. We demonstrate the robustness and repeatability of sO 2 calculation. The unprecedented significance of our vis-OCTA oximetry is the capability of excluding confounding factors from other retinal layers through 3D segmentation. In addition, the accurate vessel localization of vis-OCTA allows for measurement of capillary sO 2 , which is unattainable by fundus-based oximetry.
One limitation of this study is the lack of independent validation of our sO 2 calculation. In our previous studies on rodent retinal oximetry, a pulse oximetry was used to serve as a gold standard for arterial sO 2 close to the ONH [27,32]. sO 2 alterations due to oxygen diffusion as blood travels through the retina would limit this approach for validation of measurements in microvasculature distant from the ONH. From our calculations, we observe the clear separation of alternating feeding arterioles and draining venules, which serves as a physiologic and anatomic validation. Yet we still lack a gold standard for capillary sO 2 . Further physiologic (e.g. hypoxia challenge) or pathologic validations with known impact on oxygen levels are warranted to rigorously validate the accuracy and sensitivity of microvascular sO 2 by vis-OCTA.
We should also emphasize that all the spectral extraction was performed on structural vis-OCT signals, with vis-OCTA provided vessel localization. Because of the stringent requirements for imaging speed and SNR for vis-OCTA, the FOV was limited to 3×7.8 mm 2 at each acquisition with 6s total acquisition time at the current system characterization. This may be a limiting factor for large-scale clinical applications. Alternatively, we could use a dual-channel system with both visible and conventional near infrared (NIR) wavelengths to simultaneously acquire OCTA images [7,24]. While the NIR channel could provide rapid OCTA with a larger FOV for vessel localization, the co-registered vis-OCT structural data could provide spectral contrast for sO 2 .
To summarize, we have for the first time demonstrated vis-OCTA human retinal imaging. By achieving a 100 kHz A-line rate, the fastest acquisition speed reported so far for human retinal vis-OCT, we can acquire a single human retinal imaging data with a FOV of 3×7.8 mm 2 within 6 seconds. The angiography also enables accurate localization of microvasculature to allow us to perform human retinal oximetry in vessels less than100 µm in diameter. This technical advance lays a foundation for assessing human retinal local oxygen extraction and metabolism.