1.

Introduction

Tissue microcirculation involves the smallest vessels of the circulatory system: a network of capillaries that supply cells in the body with oxygen and nutrients and simultaneously remove waste products emanating from the cell breathing cycle. When delivery of blood or cellular uptake of oxygen is severely impaired, cell failure is soon to be found in the affected region. Anomalies in distribution at a local microcirculatory level can reveal early signs of sickness through parameters such as hemoglobin oxygen saturation level and red blood cell (RBC) tissue fraction.¹ Baseline values may reveal important physiological information, but more often the circulatory system’s ability to adapt to natural or induced stimuli is evaluated in clinical settings. Stimulus–response protocols reveal the microcirculatory system’s ability to constrict or dilate vessels and thereby regulate the blood flow to adapt to periods of ischemia or heating provocation.^2–5

Oxy- and deoxyhemoglobin, found in RBCs, are strong absorbers in the visible wavelength range in human tissue with unique spectral fingerprints. The amount of oxy- and deoxyhemoglobin may be estimated using diffuse reflectance spectroscopy (DRS) to study hemoglobin oxygen saturation and RBC tissue fraction. DRS uses a broadband light source to irradiate tissue. Photons that enter the tissue interact with structures and chromophores through scattering and absorption processes. Backscattered photons that are detected and resolved spectrally thus carry information about tissue composition. Depending on the geometry in a fiber optic probe or in an imaging setup, different photon transportation models need to be applied to untangle this information. Simplified homogeneous or layered tissue models combined with diffusion theory or Monte Carlo simulations can be used to decompose the measured DRS signal.⁶

Multispectral imaging (MSI) and hyperspectral imaging (HSI) are DRS-based methods providing 2D images containing spectral information about every pixel, creating a 3D dataset, i.e., a data cube. The acquisition of data cubes is commonly done either by sweeping through the wavelength dimension (spectral scanning) or the spatial dimensions (spatial scanning). Spatial scanning can be done by acquiring the complete spectrum for a single pixel (whiskbroom technology) or for a line of pixels (pushbroom technology), and then scanning through the area of interest. In spectral scanning, a 2D image is captured at each exposure, which involves scanning through several wavelengths.⁷ In an occlusion-release experiment, Bjorgan et al.⁸ used a pushbroom hyperspectral camera system on the volar forearm to estimate oxygen saturation levels at baseline and after 5 min of occlusion. They found a systematic difference between Monte Carlo simulations and the diffusion model in estimating oxygen saturation, emphasizing the need for an accurate model when analyzing data cubes.

MSI snapshot cameras capture the full data cube (both spatial and spectral information) in a single exposure allowing for temporal changes of the microcirculation to be monitored.⁹ The spectral information is often achieved by applying filters to individual detector elements or by dispersing the light spectrally over the detector, which limits both spectral and spatial dimensions of the data cube.¹⁰ The snapshot MSI detector used in this work has a Fabry–Perot filter array with 16 filters distributed in a recurring $4\times 4$ mosaic pattern over the sensor pixels.¹¹ Both Bauer et al.¹² and Rubins et al.¹³ used the described snapshot MSI camera to capture data during occlusion-release tests. Rubins et al. measured the palm of the hand and fitted data from 10 wavelength bands to modeled data using a two-layered diffusion model where hemoglobin oxygen saturation, RBC tissue fraction, melanin concentration, and epidermal thickness varied.^13,14 Bauer et al.¹² captured data from the back of the hand and analyzed intensity data from three spectrally corrected wavelength bands using a modified Beer–Lambert expression.

The Monte Carlo method for light transport has been considered a gold standard that allows light transport in arbitrarily complex geometries to be modeled at the expense of computational time.⁶ Therefore, we have used it in our previous work¹⁵ describing an MSI snapshot camera estimating hemoglobin oxygen saturation and RBC tissue fraction in skin microcirculation. In this paper, we have advanced the method by changing the white normalization procedure as well as the hemoglobin absorption spectrum. Moreover, the method was applied to a cohort of 24 subjects undergoing standardized vascular occlusion tests with large dynamic changes in hemoglobin oxygen saturation (arterial occlusion) and RBC tissue fraction (venous occlusion). The work aimed at extensive in-vivo validation of the MSI system’s ability to quantify hemoglobin oxygen saturation and the relative RBC tissue fraction, comparing MSI data to data co-registered with a probe-based reference system. The reference system analyzes DRS data in the same spectral range as the MSI system.^16,17

2.

Materials and Methods

2.7.

Post-processing of Monte Carlo Simulated Spectra

Post-processing of Monte Carlo simulated spectra is needed due to the substantial spectral overlap of the 16 wavelength channels of the MSI camera. This was previously described in Ewerlöf et al.¹⁵ To compare a Monte Carlo generated spectrum, $H^{\mathrm{MC}}(\lambda )$, with the measured 16 channel spectrum, the intensity, $I^{\mathrm{MC}}(n)$, corresponding to each channel, was calculated from Eq. (6), where $r_n(\lambda )$ is the camera detector sensitivity given by the manufacturer for each channel and $L(\lambda )$ is the output spectrum of the light source. The response is summarized for all wavelengths in the range of 450 to 650 nm to get the total intensity response from each channel. The procedure is illustrated in Fig. 1.

Eq. (6)

$$I^{\mathrm{MC}}(n)=\sum _{\lambda }{r}_n(\lambda )·L(\lambda )·H^{\mathrm{MC}}(\lambda ).$$

Fig. 1

An outline describing how the modeled intensities for the 16 channels (a) are calculated. For each channel, the modeled intensity is found by summarizing the channel detector sensitivity (b), the light source emission spectrum (c), and the Monte Carlo generated spectrum (d) over the wavelength range as described by Eq. (6). The Monte Carlo generated spectrum is found in the pre-simulated look-up table for a combination of model parameters.

The corresponding white normalization,$I_{\mathrm{w}}^{\mathrm{MC}}(n)$, was calculated by Eq. (7), using the generated spectrum $H_{\mathrm{w}}^{\mathrm{MC}}(\lambda )=1$, and then used in Eq. (8) to get the white normalization intensity response, $I_{\mathrm{N}}^{\mathrm{MC}}(n)$, for a Monte Carlo generated spectrum.

Eq. (7)

$$I_{\mathrm{w}}^{\mathrm{MC}}(n)=\sum _{\lambda }{r}_n(\lambda )·L(\lambda )·1,$$

Eq. (8)

$$I_{\mathrm{N}}^{\mathrm{MC}}(n)=\frac{I^{\mathrm{MC}}(n)}{I_{\mathrm{w}}^{\mathrm{MC}}(n)}.$$

3.

Results

Table 1 presents the average hemoglobin oxygen saturation and RBC tissue fraction estimated by MSI and EPOS systems during arterial and venous occlusion of all subjects during the three time intervals indicated by gray horizontal lines in Fig. 2(a). The subject used as an example in Figs. 2, 4, and 5 was chosen based on the correlation coefficient $R^2$ between MSI and EPOS estimated values of hemoglobin oxygen saturation being closest to the mean of all subjects.

Table 1

Hemoglobin oxygen saturation (SO2 ) and RBC tissue fraction (fRBC) estimated by MSI and EPOS at baseline, end of occlusion, and start of reperfusion (mean (SD) and [min max] for all included subjects). For fRBC, the second line presents MSI values calibrated to EPOS data.

Fig. 2

(a) Hemoglobin oxygen saturation (${\mathrm{S}}_{{\mathrm{O}}_2}$) during a typical arterial occlusion (MSI = bold line, EPOS = thin line). Occlusion starts at $t=300\mathrm{s}$ and ends at $t=600\mathrm{s}$. Vertical lines indicate time interval used for regression analysis (240 to 780 s). Time intervals used for Bland–Altman analysis are marked in gray for baseline (240 to 295 s), occlusion (540 to 590 s), and reperfusion (595 to 655 s). (b) Linear regression analysis of data in (a) (every 10th data point). Data during baseline (squares), occlusion (circles), reperfusion (triangles), and the fitted regression (solid line) are presented.

The changes in oxygen saturation for the chosen subject during an arterial occlusion provocation are presented in Fig. 2(a). The mean oxygen saturation level at baseline estimated by the MSI system is 41.5% in the first indicated time interval (gray horizontal lines). When occlusion is initiated, it decreases and after 5 min of occlusion, oxygen saturation levels are down to only 2%. When occlusion is released, oxygen saturation peaks at 84.7% during the hyperemic period. Oxygen saturation then decreases toward the baseline level.

Oxygen saturation estimated by the MSI system was compared to EPOS data by linear regression analysis. Figure 2(b) presents sampled data from Fig. 2(a) and the plotted linear regression. For this subject, $R^2$ is 0.964 and the slope of the line is 1.05. For the whole population of 21 included subjects, the mean value of $R^2$ is 0.958 with values ranging between 0.904 and 0.993.

Bland–Altman analysis was performed for 21 subjects at three chosen time intervals indicated in Fig. 2(a). The result is presented in Fig. 3. MSI and EPOS methods give similar results for low oxygen saturation at the end of occlusion. The mean bias is 0.3% with upper and lower limits of 3.6% and $-4.2\%$ deviation, respectively. Baseline measurements show a larger difference with a mean bias of 13.2% with upper and lower limits of 4.3% and $-30.7\%$ deviation. EPOS values are generally considerably higher than MSI values. Higher EPOS values are also found during reperfusion but not to the same extent. Mean bias is 5.6% with upper and lower limits of 4.4% and $-15.9\%$ deviation, respectively.

Fig. 3

Agreement of hemoglobin oxygen saturation $(S_{{\mathrm{O}}_2})$ between values estimated by MSI and EPOS from Bland–Altman analysis of 21 subjects at three separate time intervals [see Fig. 2(a)] during arterial occlusion: baseline (squares/dashed), end of occlusion (circles/solid lines), and reperfusion (triangles/dash-dotted lines). The mean difference for each phase is represented by thin lines and the difference $\pm 1.96$ * standard deviation (95% prediction interval) by bold lines.

Figure 4 shows spatially resolved oxygen saturation images calculated at three different temporal points for the same subject as in Fig. 2. At baseline and occlusion, the last timepoint in the intervals indicated in Fig. 2(a) was selected, whereas the timepoint with the peak oxygen saturation was selected within the reperfusion interval. The three intensity data cubes acquired by the MSI system were analyzed pixel by pixel, displaying the oxygen saturation of the lower arm.

Fig. 4

Spatially resolved oxygen saturation estimated by MSI. For baseline (a) and occlusion (b), the last timepoints of the intervals indicated in Fig. 2(a) were selected, whereas the timepoint with the peak oxygen saturation in the interval was selected for reperfusion (c). The ROI used for averaging data presented in Fig. 2 is outlined by the white square.

RBC tissue fraction data estimated during venous occlusion for the same subject as in Fig. 2 are presented in Fig. 5(a). Levels are stable during baseline but increase rapidly when the arm is occluded. During occlusion, the increase continues but at a slower rate. Right after occlusion release, levels drop and soon reach baseline level again. Estimated EPOS RBC tissue fraction values are 0.38% at baseline and 0.79% at the end of occlusion. MSI data are calibrated as described in Sec. 2.9 prior to comparison with EPOS data.

Fig. 5

(a) RBC tissue fraction ($f_{\mathrm{RBC}}$) during a typical venous occlusion (calibrated MSI = bold line, EPOS = thin line). For MSI data calibration, see Sec. 2.9. Occlusion starts at $t=300\mathrm{s}$ and ends at $t=600\mathrm{s}$. Vertical lines indicate time interval used for regression analysis (240 to 780 s). Time intervals used for Bland–Altman analysis are marked in gray for baseline (240 to 295 s) and occlusion (540 to 590 s). (b) Linear regression analysis of data in (a) (every 10th data point). Data during baseline (squares), occlusion (circles), reperfusion (triangles), and the fitted regression (solid line) are presented.

Linear regression analysis was performed on RBC tissue fraction values at venous occlusion for all 20 included subjects. Figure 5(b) presents regression analysis of data in Fig. 5(a). For this subject, $R^2$ is 0.982 and the slope of the line is 0.61. For all included subjects, the mean value of $R^2$ is 0.925 with values ranging between 0.717 and 0.984. $R^2>0.9$ for 75% of the subjects.

Results from Bland–Altman analysis of 20 subjects are presented in Fig. 6 for two time intervals indicated in Fig. 5(a). After normalization, RBC tissue fraction is underestimated by MSI measurements at baseline with mean bias $-0.18$ and upper and lower limits 0.014 and $-0.37$, respectively, compared with EPOS. At the end of venous occlusion, RBC tissue fraction is underestimated by MSI compared to EPOS with mean bias 0.14 and upper and lower limits 0.49 and $-0.21$, respectively.

Fig. 6

Agreement of RBC tissue fraction ($f_{\mathrm{RBC}}$) between values estimated by MSI (calibrated) and EPOS from Bland–Altman analysis of 20 subjects at two separate time intervals [see Fig. 5 (a)] during venous occlusion: baseline (squares/dashed lines) and during end of occlusion (circles/solid lines). The mean difference for each phase is represented by thin lines and the difference $\pm 1.96$ * standard deviation (95% prediction interval) by bold lines.

The ratios of estimated RBC tissue fraction between end of occlusion and baseline, by MSI and EPOS, are presented in Table 2 for both arterial and venous occlusion. The ratio was significantly larger when measuring with MSI ($p<0.001$, arterial and venous occlusion), and most distinct during venous occlusion when change in RBC tissue fraction is apparent.

Estimated average maximal photon penetration depth at the two wavelengths 475 and 630 nm is 0.43 and 0.79 mm with EPOS (each value is an average of the result for the DRS source detector fiber at 0.40 and 1.20 mm distances), respectively. Similar MSI values are 0.45 and 1.79 mm, respectively.

Table 2

Ratio of estimated RBC tissue fraction between end occlusion and at baseline for MSI and EPOS during arterial and venous occlusion (mean and (SD) for all included subjects).

4.

Discussion

In this study, we have advanced and evaluated an algorithm estimating microvascular hemoglobin oxygen saturation and RBC tissue fraction from collected snapshot MSI data. The algorithm, based on inverse Monte Carlo modelling, uses six parameters to model light transport in tissue, with three of the parameters assigned fixed values and three varying within a specified range relevant to skin tissue. Measurements during venous and arterial forearm occlusions of 24 subjects show that assessment of hemoglobin oxygen saturation and RBC tissue fraction with our MSI system and algorithm is highly correlated with those measurements using a probe-based reference system. However, there are systematic differences that might be explained by differences in sampling volume.

When synchronizing in time between MSI and EPOS, reactions to occlusion and reperfusion seem to be concurrent in both methods. Brachial vascular occlusion affects the whole arm distally, and therefore the two measurement sites (MSI and EPOS) are influenced similarly. Spatial differences in tissue compounds between the sites, for example, presence of shallow or large vessels or other structures, may impact the result at an individual level but are averaged out within a population.

Hemoglobin oxygen saturation values are estimated and verified through a linear regression analysis showing a mean correlation coefficient $R^2=0.956$, compared to the EPOS reference system. Applying Bland–Altman analysis of MSI and EPOS shows that agreement is best at the end of occlusion when oxygenation is low. There is an offset between MSI and EPOS measurements during baseline, where MSI shows a lower hemoglobin oxygen saturation.

Since MSI RBC tissue fraction values are relative in contrast to values measured by EPOS, MSI values are calibrated to those from EPOS before comparison. Applying linear regression analysis shows an $R^2=0.927$, whereas the Bland–Altman analysis reveals that MSI underestimates at baseline and similarly overestimates the end of occlusion values compared to EPOS RBC tissue fraction results.

The additional Monte Carlo simulation on penetration depth shows that, for longer wavelengths where light absorption of hemoglobin is low, the MSI system has a sampling volume reaching substantially deeper into tissue (1.79 mm) compared to EPOS (0.79 mm). The most superficial veins are found at a depth around 0.75 mm in forearm tissue,²⁶ which suggests that they are mainly undetectable to the EPOS system. The systematic differences between the MSI and EPOS values may be explained by this difference in sampling volumes, effectively making the MSI measurements more influenced by venous blood.

During end of arterial occlusion, both MSI and EPOS sample a vascular bed with essentially deoxygenated hemoglobin. Similarly, both methods sample a vascular bed with essentially oxygenated blood soon after release of the occlusion due to a short capillary transit time and a low oxygen extraction during hyperemia, caused by a high blood flow speed. However, at baseline, the difference in oxygenation between vascular plexa is more pronounced, which makes the different sampling volumes of the two modalities affect their oxygen measures more strongly.

Sampling volume differences can also explain the larger change in RBC tissue fraction measured by MSI than EPOS when comparing baseline and end of occlusion data in Table 2. The observation holds both for arterial and venous occlusion but is more expressed for the latter. Deeper lying venous vessels have a high capacitance and the RBC tissue fraction in these vessels is increased when blood is accumulated during occlusion. The MSI with its deeper sampling volume will sample more of the venous blood than EPOS, which would explain the difference in estimation of the blood volume.

During venous occlusion, RBC tissue fraction increases gradually throughout the length of the occlusion period as expected. When analyzing RBC tissue fraction levels during arterial occlusion (data not presented), RBC tissue fraction similarly increases during cuff inflation but also changes after the cuff pressure reach levels well above the arterial pressure. This could be explained by a redistribution of blood driven by the initial pressure difference between arteries and veins, causing blood to flow from deep arteries positioned well below our sampling depth to veins located deep but visible to our measurement systems.

It is possible to apply the described algorithm on individual pixels of the MSI data cube to produce spatially resolved results, as illustrated in Fig. 4. In this study, MSI data from a $2\times 2\mathrm{cm}$ large area was averaged, analyzed, and compared to data from the point-wise measuring EPOS probe. The example in Fig. 4 shows small spatial variations in oxygen saturation within the ROI, which indicates that averaging has a limited impact on the result. The spatially averaged MSI data in comparison to the EPOS data may show less conformance due to different sampling sites. However, for the whole study, MSI and EPOS values agreed well, especially for hemoglobin oxygen saturation where the $R^2$-value is $>0.90$ for all subjects.

Other groups have used the snapshot MSI camera from XIMEA to measure skin tissue oxygen saturation. Bauer et al.¹² describe a method for spectral correction of the XIMEA camera characteristics, where unwanted spectral peak sensitivity is decreased. This is essential when using inverse algorithms that do not fully account for the complete spectral sensitivity of each pixel, but instead analyze the data based on scattering and absorption values for a single wavelength (i.e., the peak wavelength). Our approach does not suffer from this limitation as the complete spectral behavior of both skin tissue and detector characteristics are accounted for. However, this is done at the expense of a computationally demanding algorithm based on Monte Carlo simulations, making it unsuitable for real-time analysis of complete images. Bauer et al. applied their algorithm to estimate relative changes in blood oxygen saturation on the back of the hand, using only 3 out of 16 available channels. The relatively low acquisition speed (1 fps) allowed values to be calculated in real time. To evaluate the system dynamically, they used arterial occlusion of the upper arm. The results are compared with a standardized probe-based system that measures in the near-infrared (NIR) regime²⁷ and sampling volume for the two systems differ considerably.¹² Bauer et al. found differences in oxygenation during occlusion like we did, but it is not possible to compare estimated oxygen saturation values, as all presented values are normalized to the range [0 1] and compare only shapes of the curves.

Rubins et al. used a $k$-means segmentation algorithm and a two-layered optical diffuse model to relate intensities measured from palm side of the hand by the XIMEA camera to skin oxygen saturation levels. Their protocol involved upper arm occlusion to 150 mmHg for 10 min.¹³ The results cannot be compared to our values as values are not quantitative or compared to a reference method. The cuff pressure was considerably lower than the pressure used in our protocol and might not have given a similar response. Furthermore, the dynamic response in the reperfusion phase was limited to the same level as oxygen saturation values estimated during baseline (100%).¹³

The sensitivity to the hemoglobin absorption spectra is influenced by key wavelengths and an analysis algorithm may be improved by a wise selection or combination of channels.^22,28 However, these sensors are unique, and spectral properties, such as peak wavelength, vary between specimens. Therefore, when designing algorithms that are expected to work not only for a single camera, fixed pre-selected wavelengths or channels are not a viable option. In this work, we present an optional strategy that addresses this. In general, our approach appears to work well. Some residuals are, however, present in our non-linear optimization algorithm when fitting simulated data to measured data, which might explain some of the differences found between the two compared modalities. The origin of these residuals can most likely be found in the spectral sensitivity provided by the manufacturer. Preliminary data indicate that these spectral sensitivities might not be exact when adding an optical lens to the sensor, which calls for further calibration measurements.

5.

Conclusion

We have demonstrated that an inverse Monte Carlo algorithm, based on a two-layer skin model with three free parameters, can accurately and consistently model spectra captured using an MSI snapshot camera with complex filter characteristics. Data acquired from forearm skin during arterial and venous occlusion show that the hemoglobin oxygen saturation and RBC tissue fraction estimated from MSI data have a high correlation with values co-registered with the validated probe-based EPOS system. Some systematic differences were found, but these may be explained by a larger sampling volume for the MSI system at longer wavelengths where tissue absorption is low.