2.1.

Wavelength Selection for Dual Wavelength Retinal Oximetry

Oxygenation is typically assessed by imaging the retinal vessels at different wavelengths using multispectral imaging techniques, where the choice of wavelengths will affect the achievable measurement accuracy. In an earlier publication, we provided a detailed explanation and calculation of retinal oximetry with and without the pigment packing effect.²⁶ Here, we present a summary of the calculations but now without assuming the use of an isosbestic wavelength for oximetry, as was done in our previous paper. Furthermore, we will account for the maximum permissible power that can be used for retinal illumination and the quantum efficiency of the detector in the calculations.

Let $I({\overline{x}}_b,\lambda )$ and $I({\overline{x}}_t,\lambda )$) be the recorded intensities in the retinal image at the center of the blood vessel and adjacent tissue location, respectively, at a wavelength $\lambda$ and for equal illumination intensity at both locations. The relative optical density ${\mathrm{OD}}_{\lambda }$ of the blood vessel location compared to the surrounding tissue can be written as follows:

We can then write the saturation (i.e., the fraction of the oxygenated hemoglobin concentration to the total concentration of hemoglobin) as follows:²⁶

Fig. 1

(a)–(e) In silico analysis of saturation error $\mathrm{\Delta }S$ for different saturation levels and wavelength combinations for ${\lambda }_1,{\lambda }_2\in$ (450 to 650) nm. The standard error on the mean recorded intensity values ($\mathrm{\Delta }I/I$) was assumed to be 1%. All values of $\mathrm{\Delta }S>0.05$ were set to 0.05.

Figure 2(a) shows an ensemble $\mathrm{\Delta }S$, where weights of 0.4, 0.6, 0.8, 1.0, and 1.0 were applied to the five different saturation graphs $S=0.00$, 0.25, 0.50, 0.75, and 1.00 in Fig. 1, respectively. The weights were chosen to favor the more physiologically relevant, health range of oxygen saturation.³¹ Figure 2(b) shows the difference in calculated saturation with and without the pigment packing effect defined as the saturation offset, $S_{\text{offset}}$ ($S_{\text{offset}}=S_{{\lambda }_1|{\lambda }_2}-S_{{\lambda }_1|{\lambda }_2,pp}$) due to the pigment packing effect when the same weights were applied. Figure 2(c) shows the addition of the saturation offset and the saturation error for wavelengths ${\lambda }_1,{\lambda }_2\in [450650]$, respectively. Finally, Fig. 2(d) shows Fig. 2(c) corrected for the wavelength-dependent maximum permissible exposure (MPE) prescribed for 30,000 s of continuous single-point exposure³² (see Fig. 15 in Sec. 5) and responsivity of the detector across the wavelengths.³³ The $\mathrm{\Delta }I/I$ value was 1% for the MPE and responsivity at 600 nm and increases linearly for decreasing MPE and responsivity as a function of the wavelength to get Fig. 2(d). For example, at 460 nm, the MPE normalized on the MPE for $\lambda \ge 500\mathrm{nm}$ was 0.1 and the responsivity normalized on the responsivity at 600 nm was 0.7 and hence, the $\mathrm{\Delta }I/I$ was adjusted to 14.3%. Based on Fig. 2(d), 498- and 594-nm wavelengths were chosen for oximetry estimation.

Fig. 2

In silico analysis of saturation error and saturation offset. (a) Saturation error $\mathrm{\Delta }S$ for different wavelength combinations for ${\lambda }_1,{\lambda }_2\in$ (450 to 650) nm. The standard error on the mean recorded intensity values ($\mathrm{\Delta }I/I$) was assumed to be 1%. All values of $\mathrm{\Delta }S>0.05$ are shown as 0.05. (b) Calculated saturation difference with and without pigment packing effect shown as an offset ($S_{\text{offset}}=S_{{\lambda }_1|{\lambda }_2}-S_{{\lambda }_1|{\lambda }_2,pp}$) for a $50\text{-}\mu \mathrm{m}$ diameter blood vessel. Values of $S_{\text{offset}}>0.1$ are shown as 0.1. (c) Addition of the saturation offset and the saturation error for different wavelength combinations. The color bar has a range from 0.05 to 0.15. (d) Addition of saturation offset and saturation error corrected for the MPE prescribed for 30,000 s of continuous single-point exposure (see Sec. 5), and quantum efficiency of the detector across the wavelengths.

2.2.

System Design

The scanning laser oximeter setup was an adaptation of our previous design,²⁶ with the aim to optimize the SNR of our imaging system. A schematic layout of the system is shown in Fig. 3. The light from an SC source (EXU-6, NKT Photonics A/S, Birkerød, Denmark) was filtered using a tunable filter (Select, NKT Photonics A/S). A 90:10 fused single-mode 460-HP fiber coupler (TW560R2A2, Thorlabs GmbH, Germany) was used to send 10% of the light into detection module II for continuous laser power monitoring and balanced detection (see Sec. 2.3). The core of the double-clad fiber (DCF) in port S of the DCFC^34,35 was excited using the 90% port of the splitter. About 94% (transmission measured at 550 nm) of this light reached the port A of the DCFC. A reflective fiber collimator (RC02APC-P01, Thorlabs GmbH) was used to collimate the fiber output from port A into a 2.2 mm beam. This beam size was optimum for two reasons: (i) The beam size was small enough that the beam is not clipped in the pupil of the subject even with the expected pupil constriction due to bright visible radiation. (ii) This beam size was close to the optimum pupil size for lateral resolution.³⁶ A dichroic mirror (D3, long pass cutoff 700 nm) was used to prevent any residual pump light of the SC from entering the rest of the system. A custom-made dispersion compensating lens (DCL) (Shanghai optics Inc., H-FK61 and K4A, $f=1800\mathrm{mm}$) was placed in the common beam path to compensate for the longitudinal focal shift arising from the chromatic aberration of the human eye.³⁷ A combination of a 2.56-kHz resonant scanner (5.12 kHz line rate by utilizing both sides of scanner sweep) (EPOC., Glendale, New York, United States), and a 10-Hz galvanometer mirror (Cambridge Technology, Bedford, Massachusetts, United States) placed close to each other were used to achieve an imaging throughput of $5.12\text{million}\text{pixels}/\mathrm{s}$ ($10\text{frames}/\mathrm{s}$ $512\text{lines}/\text{frame}$ $1000\text{pixels}/\text{line}$). A 1:1 telescope consisting of achromatic doublet pairs was used to relay a point approximately in the middle of the two scanning mirrors onto the pupil of the eye. The reflected and scattered light from the retina was then collected using the inner cladding of the DCF in port A. About 63% (transmission measured at 550 nm) of the collected light is then coupled into the multimode fiber. The fiber cross-section of the DCF and the MMF in the DCF coupler is also shown in Fig. 3.

Fig. 3

Schematic sketch of the scanning laser oximeter showing the various components. The abbreviated components are described in the legend. The light from port A of the DCFC was collimated using RC into a 2.2 mm beam. The reflected signal from the retina was collected using the cladding of the DCF with a diameter of $105\mu \mathrm{m}$ and a numerical aperture (NA) of 0.22. This light is coupled on to the MMF with a diameter of $200\mu \mathrm{m}$ and an NA of 0.22. Thus, the illumination pinhole size is $9\mu \mathrm{m}$ and the detection pinhole size is $105\mu \mathrm{m}$. The maximum collection beam diameter of RC was 8 mm. However, the limiting aperture in the optical setup is the pupil of the subject.

The detection modules consist of a collimating lens (L1 and L4, $f=18\mathrm{mm}$), a dichroic mirror (D1 and D2, long pass cutoff 532 nm), and two focusing lenses (L2, L3, L5, and L6, $f=25\mathrm{mm}$) focusing the spectrally separated light on the active area of avalanche photodiodes (APD) (APD410A2, Thorlabs GmbH). Two clean-up filters (CF1, CF3: bandpass $470\pm 50\mathrm{nm}$, CF2, CF4: long pass 532 nm) were placed just before the focusing lenses to avoid any contribution of light from an unintended wavelength. The detectors in both the detection modules were connected to a 12-bit digitizer (NI-PCI6115, 10 million samples/s/channel, National Instruments, Austin, Texas). The signals from the four detection modules were sampled using four parallel channels of this digitizer running on a single clock. Port B of the DCF coupler was placed in a medium, whose index was matching the core index of the MMF to suppress the back reflection from the exit surface of the fiber faucet from entering the detection module I.

2.3.

Balanced Detection to Increase the Signal-to-Noise Ratio

Spectrally filtering an SC light source provides an excellent wavelength-flexible alternative to the existing approaches of multispectral imaging. However, in an SC light source such as the EXU-6 used in the setup shown in Fig. 3, the generation of a large optical bandwidth through nonlinear processes^38,39 results in high RIN compared to traditional sources, such as LEDs, lasers, and superluminescent diodes. The fundamental noise limitation of an SC source is described well by Corwin et al.⁴⁰ The pulse-to-pulse intensity variations in the SC source amounts to the RIN. In principle, RIN can be suppressed by averaging more pulses within the integration time of the detector. The EXU-6 has a repetition rate of 78 MHz, and given the short integration times of our SLO, ($\sim 380\mathrm{ns}$) only 30 pulses were averaged within an integration time resulting in a significant RIN component in the total noise figure. The total noise power of the system can be written as follows:

Fig. 4

Noise performance of the SC source shown by plotting the SNR for different levels of mean signal in the photodetector for two wavelengths, 500 and 600 nm. The graph shows the theoretical shot noise limited SNR (solid black line) and the expected SNR due to the additional contribution of thermal and quantization noise to the shot noise (dotted black line). The measured SNR for two different wavelengths 500 and 600 nm is also shown (dashed green and red line, respectively). The SNR deteriorates for higher photocurrents because RIN noise scales as the square of detected photons in the photodetector. By using balanced detection, an improvement in SNR of $\sim 5\mathrm{dB}$ was achieved for these wavelengths (dotted green and red lines).

2.4.

Wavelength Sweep Hyperspectral Imaging

The acousto-optic tunable filter (Select, NKT Photonics A/S) used in the system (Fig. 3) allows to perform a wavelength sweep for the desired wavelength ranges between 450 and 650 nm and in desired steps (minimum step size 0.1 nm, maximum sweep speed of 10 wavelengths/s). We performed a wavelength sweep from 485 to 608 nm with the SLO with a step size in the wavelength of 3 nm and sweep speed of 1 step/s. Given that the imaging speed of our SLO is 10 fps, this enabled acquiring high-quality retinal images at a rate of 10 images/wavelength within 42 s of imaging. For these measurements, the dichroic filters D1, D2 and the clean-up filters CF2 and CF4 (see Fig. 3) were removed and only one channel was used in each detection unit.

2.5.

In Vivo Human Measurements

The clinical pilot study obeyed the principles of the “Declaration of Helsinki” and was approved for human use by the medical ethical review board at the Amsterdam Universitair Medische Centra (VUmc location), Amsterdam. Tropicamide 0.5% wt./vol. drops were administered for pupil dilation during imaging. The optical power used for measuring human subjects was in agreement with the MPE established by the latest IEC standard 60825-1.³² The laser safety considerations are detailed in Sec. 5. Different field-of-views (FoVs) could be implemented by changing the voltages provided to the scanners. Background measurements were taken prior to each imaging session and were subtracted from SLO images to remove artifacts, such as lens reflections. In order to utilize the full dynamic range of the detector, strong lens reflections from the last ophthalmic lens in the system were allowed to saturate the detector. All other stray reflections that did not saturate the detector were removed using the background measurements.

2.6.

Retinal Vessel Segmentation and Oxygenation Map

The SLO images were background corrected by performing a reference measurement before imaging in vivo. The background subtracted images were corrected for the sinusoidal scanning of the resonant scanner into a single image. The retinal oxygenation was constructed using the flow diagram shown in Fig. 5. First, several linear intensity images were registered using the scale-invariant feature transform method.⁴¹ After correcting the images for translation, rotation, and affine transformations, an averaged image was constructed from the registered data. In order to segment and extract the vessels in the retinal image, the averaged image $\mathrm{Im}(x,y)$ was normalized as ${\mathrm{Im}}_N(x,y)=\frac{\mathrm{Im}(x,y)-\min [\mathrm{Im}(x,y)]}{\max [\mathrm{Im}(x,y)]-\min [\mathrm{Im}(x,y)]}$, where min and max are defining the minimum and the maximum values in the registered and averaged image. Under the assumption that the SLO images have an approximate piecewise constant structure, the images were denoised using a total variation (TV) algorithm in order to preserve the vessel edges. Subsequently, contrast enhancement was applied to determine the blood vessel boundaries correctly. As has been shown by Davidoiu et al.,⁴² a regularization TV model using a specialized fixed-point algorithm and isotropic TV norm improved the 2-D vasculature network segmentation for ex-vivo heart data. In our case, the regularization parameter in the TV algorithm has been fixed and chosen to be equal to the standard deviation of the normalized image and the convergence stopping parameter value has been set to 0.01. The contrast enhancement of the denoised image was performed using a method called contrast-limited adaptive histogram equalization.⁴³ Finally, the blood vessels detection and diameter calculation was performed using wavelets and edge location refinement.⁴⁴ The extracted vessel boundary coordinates were exported to MATLAB and were applied to the averaged image Im $(x,y)$. The extraction of the blood and tissue intensity values was done using the method described by Damodaran et al.²⁶

Fig. 5

Pipeline for vessel segmentation and oxygenation extraction from retinal images.

3.1.

Technical Aspects Regarding Multispectral SLO with an SC Source

3.1.1.

SNR improvement by balanced detection

Figure 6 shows the comparison between in vivo retinal images with and without balanced detection in three different volunteers with three different wavelengths. These images clearly show the improvement in image quality due to balanced detection. As described in Sec. 2.3, the RIN of the SC source was mitigated by using a balanced detection scheme. This resulted in an SNR improvement of 5 dB.

Fig. 6

(a)–(e) Comparison of images with and without balanced detection—unaveraged retinal images from three different volunteers show the improvement in image quality due to balanced detection. The red arrows show the blood vessels that become visible only due to balanced detection. The green arrows point to the lens reflection artifact that saturated the detector and was not removed due to background subtraction. Scale bar represents $300\mu \mathrm{m}$ in the retina.

3.1.2.

Compensation for chromatic aberrations of the eye

Chromatic aberrations in the human eye have been studied previously.^45,46 The longitudinal chromatic aberration (LCA) is a potential source of error in oximetry due to the dispersion in the human eye.³⁷ LCA causes the spatial location of each pixel in the multispectral image to be in a slightly different axial plane for different wavelengths. This effect can lead up to a $220\mu \mathrm{m}$ longitudinal focal shift between the 498- and 594-nm wavelengths.⁴⁷ This effect can be corrected, as has been shown previously by others,^48,49 by using a DCL in the beam path.

Images acquired at 500 and 600 nm without the dispersion compensation lens (Fig. 7) in the right eye of a healthy human volunteer demonstrate that chromatic aberrations affect the multiwavelength SLO images; note the difference in the appearance of the foveal pit in the 500 nm and 600 nm images. This can be explained by a focal position that is at the depth of the foveal pit for the 600 nm image, and the 500 nm image has a shallower focus, as evidenced by the dark appearance of the foveal pit due to confocal gating. This focal separation between the different wavelengths was compensated by a dispersion compensation lens (Fig. 3) that resulted in the images [Figs. 7(c) and 7(d)] having a more uniform appearance of the foveal pit between the 500 and 600 nm channels. These images are similar to LCA-compensated images published by LaRocca et al.⁴⁹

Fig. 7

Images from the right eye of a human volunteer show the effect of LCA in the absence or presence of a dispersion compensation lens. The foveal pit (red arrows) appears more uniform in the images with dispersion compensation [panels (c) and (d)] compared to the images without dispersion compensation [panels (a) and (b)]. Scale bar represents $300\mu \mathrm{m}$ in the retina.

The transverse chromatic aberrations between the wavelengths were compensated in postprocessing of the images by applying a transformation metric, which aligns the images of different wavelengths to a chosen wavelength.

3.1.3.

Effect of the size of the illumination and collection apertures of the DCF coupler

We determined the effect of using a DCF with different core and inner cladding diameters on throughput and sharpness of the image. DCF couplers with the same multimode fibers, but different DCFs (different core and inner cladding diameters) were used. Figure 8(a) shows an unaveraged SLO image of right eye of a healthy male volunteer imaged with a DCF having a core diameter of $2.3\mu \mathrm{m}$ (single mode at 430 nm) and an inner cladding of compared with a DCF having a core diameter of $9\mu \mathrm{m}$ and inner cladding diameter of $105\mu \mathrm{m}$ [Fig. 8(b)]. The figures show that while the $2.3\mu \mathrm{m}/15\mu \mathrm{m}$ DCF gives images with sharper appearance, this comes at the cost of throughput, as indicated by the intensity scale bars. For subdiffuse measurements with higher throughput requirements for fast quantitative imaging, the $9\mu \mathrm{m}/105\mu \mathrm{m}$ DCF is better suited since the amount of averaging required to achieve an SNR of 20 dB is smaller than for the $2.3\mu \mathrm{m}/15\mu \mathrm{m}$ DCF.

Fig. 8

Unaveraged retinal SLO images (single frame each) from the described design in Sec. 2.2 with different DCFs. (a) DCF with a core $2.3\mu \mathrm{m}$ diameter and an inner cladding of $15\mu \mathrm{m}$. (b) DCF with a core $9\mu \mathrm{m}$ diameter and an inner cladding of $105\mu \mathrm{m}$. The cross-section of the DCF is shown in the bottom diagrams. Scale bar is $300\mu \mathrm{m}$ in the retina.

3.2.

In Vivo Two Wavelength Oximetry

Retinal oximetry was demonstrated in two healthy volunteers using the optimal two wavelengths, 498 and 594 nm. Figures 9(a) and 9(b) show the retinal images at 498 and 594 nm, respectively. After vessel segmentation and analysis, as described in Sec. 2.6, an oxygen saturation map of the blood vessels was created and overlaid on the 594 nm image, as shown in Fig. 9(c). Figure 9(d) shows an image taken with the Oxymap Tx for comparison. Similar measurements were made in volunteer 2, as shown in Fig. 10. The smallest vessels for which oxygen values are provided are $40\mu \mathrm{m}$ in diameter.

Fig. 9

SLO images taken from a healthy human volunteer at two wavelengths, (a) 498 nm and (b) 594 nm. Ten frames were registered and averaged to improve the SNR of the images. (c) The estimated oximetry results are shown. (d) The image acquired using Oxymap Tx ehf is shown for comparison. Scale bar represents $300\mu \mathrm{m}$ in the retina.

Fig. 10

(a) Estimated oxygenation values in volunteer 2 and (b) corresponding image by Oxymap Tx. Scale bar represents $300\mu \mathrm{m}$ in the retina.

To analyze the repeatability of the oxygenation measurements, volunteer 1 was imaged three times with 8-min interval in between the measurements. The average oxygenation in four vessels (vessels 1 to 3: arteries; vessel 4: vein) is shown in Fig. 11. The calculated oxygenation values in these vessels reproduce very well, despite the difference in positioning of the eye between the three imaging sessions, resulting in different overall intensity levels and differences in focus depth, as evidenced by the different appearance of the foveal pit across the three sessions.

Fig. 11

Retinal imaging was performed in volunteer 1 with an interval of 8 min to study the repeatability of the measurements. (a)–(c) Measurements made at 0, 8, and 16 min in volunteer 1 with the calculated oximetry for three large arteries and a vein. Scale bar represents $300\mu \mathrm{m}$. The numerical values are plotted in (d).

3.3.

Wavelength Sweep Hyperspectral Imaging

Similar to other SLO systems, the system described in Fig. 3 can image different sections of the retina at high frame rates, but the SC source used in our system also has the advantage of imaging at any desired wavelength and to sweep wavelengths in selected ranges and steps in the 484 to 700 nm range. Figure 12 shows the in vivo result of a wavelength sweep from 485 to 608 nm performed in volunteer 2. The measurement took 42 s to complete, and each wavelength image shown in the figure was registered to the first frame of that wavelength and averaged in postprocessing. As a result of the balanced detection and well-fixated eye of the volunteer, the images exhibit good contrast.

Fig 12

Wavelength sweep from 485 to 608 nm.

Six blood vessels providing a wide range of vessel diameters were selected from the spectral cube, and the OD values were obtained for these vessels after performing the vessel segmentation in individual images. The weighted mean and the standard deviation (weights = 1/standard error) were obtained for these vessels and Eq. (1) (where $G$ was assumed to be 0 and the blood hemoglobin concentration was assumed to be $15\mathrm{g}/\mathrm{dL}$) was used as the model to perform a weighted fit the OD values in order to extract the $⟨L_{\mathrm{eff}}⟩$ and $S$ values, as shown in Fig. 13. The $S$ values of these vessels are similar to the values obtained by dual-wavelength imaging of the same volunteer (Fig. 10) and oxygen saturation maps obtained with Oxymap Tx ehf. The errors on the fitted parameters were calculated from the diagonals of the covariance matrix.⁵⁰ From the fitted values of $⟨L_{\mathrm{eff}}⟩$, the relationship between $⟨L_{\mathrm{eff}}⟩$ and $D_{\mathrm{ves}}$ for these vessels ($D_{\mathrm{ves}}\ge 70\mu \mathrm{m}$) could be established, and this relationship is plotted in Fig. 14. The relation between $⟨L_{\mathrm{eff}}⟩$ and $D_{\mathrm{ves}}$ was found to be fitted well by a linear function, $⟨L_{\mathrm{eff}}⟩(\mu \mathrm{m})=(0.46·D_{\mathrm{ves}}+5.2)\pm 6.0(\mu \mathrm{m})$, where $⟨L_{\mathrm{eff}}⟩$ and $D_{\mathrm{ves}}$ are given in micrometers. From this relationship, we conclude that our subdiffuse SLO imaging scheme, where a small illumination spot is used in combination with a large collection aperture, works in a light transport regime between a fundus camera (where $⟨L_{\mathrm{eff}}⟩=D_{\mathrm{ves}}$) and a truly confocal SLO, where $⟨L_{\mathrm{eff}}⟩=⟨L_{\mathrm{bs}}⟩$ with $⟨L_{\mathrm{bs}}⟩$ is the backscatter path length, which is not expected to have a linear dependence on $D_{\mathrm{ves}}$. We expect that the exact relation between $⟨L_{\mathrm{eff}}⟩$ and $D_{\mathrm{ves}}$ depends on the details of the measurement geometry and will, for example, be different for different DCF cladding sizes or, equivalently, for different confocal aperture sizes. Furthermore, we have here in our fits assumed $⟨L_{\mathrm{eff}}⟩$ to be wavelength independent, whereas it is expected that $⟨L_{\mathrm{eff}}⟩$ depends on the scattering and absorption coefficients of blood and is, therefore, in reality, wavelength dependent. This is a topic for future investigations.

Fig. 13

Plot of OD as a function of wavelength for six blood vessels in the right eye of a healthy volunteer.

Fig. 14

Linear regression fit of the effective path length $⟨L_{\mathrm{eff}}⟩$ as a function of the blood vessel diameter $D_{\mathrm{ves}}$.

Fig. 15

AEL for different wavelengths in 450 to 650 nm considering a collimated static beam for 30,000 s.

One notable advantage of the wavelength sweep using the SLO is the availability of the spectral properties of the entire retina for multiple wavelengths. This will be in the future permit expansion of the diagnostic capabilities of an SLO by providing spectral fingerprints of additional chromophores such as carotenoids. However, the imaging speed of the wavelength sweep should be dramatically improved before the clinical implementation of wavelength sweeping would become feasible.