2.1.

Sample Preparation

Blood samples were drawn through cardiac puncture from C57BL/6 female mice with ages ranging from 6 to 8 weeks, while they were kept under deep anesthesia. All procedures were conducted in accordance with the Guidelines for the Care and Use of Laboratory Animals²¹ and were approved by the institutional animal ethics committee. Euthanasia was performed by overdose of chemical anesthetics, and death was confirmed by cervical dislocation.

After blood collection, the samples were heparinized ($166\mathrm{IU}/\mathrm{mL}$) to prevent coagulation and $90\mu \mathrm{L}$ aliquots were distributed in triplicate over a 96-well microtiter plate. Four groups were devised by adding glucose solutions with different concentrations. The solutions were prepared with phosphate-buffered saline (1 M, pH 7.4) with increasing concentrations of glucose [D - (+) - Glucose, Sigma-Aldrich] between 160 and $310\mathrm{mg}/\mathrm{dL}$, with addition steps of $50\mathrm{mg}/\mathrm{dL}$. Aliquots of $10\mu \mathrm{L}$ of these solutions were added to the wells previously filled with heparinized blood, resulting in a final volume of $100\mu \mathrm{L}$ per well.

Intending to study the effect of a high multiple scattering component of the signal attenuation, as is the case when probing through biological tissue,^22,23 a second experiment was performed with a new round of blood collection. Aiming to minimize possible errors during preparation, we decided to work with only one glucose solution ($530\mathrm{mg}/\mathrm{dL}$ in concentration) for the new samples. By simple dilutions, several glucose concentrations were obtained from this high-concentrate solution. This time, glucose was diluted in an isotonic saline in order to obtain the intended higher multiple scattering effect, as previously described by Popescu et al.²⁴ These glucose solutions were added to heparinized blood, again placed in a 96-well microtiter plate, and the final samples were composed of $40\mu \mathrm{L}$ of heparinized blood and $60\mu \mathrm{L}$ saline-glucose solution, totaling a final volume of $100\mu \mathrm{L}$ per sample.

For both sets of experiments, the samples were carefully handled and maintained under controlled temperature (18°C) to avoid blood degradation and hemolysis during the experimental course. Before each data collection (blood glucose assessment, OCT measurements, and speckle analysis), each sample was slowly homogenized per the standard with a micropipette in order to prevent sedimentation, as OCT has been shown to be sensitive to effects caused by blood aggregation and sedimentation.²⁵

2.2.

Blood Glucose Assessment

For reference, blood glucose levels were assessed in duplicate by a portable glucometer (OneTouch^® Ultra^®, Johnson & Johnson Medical Devices & Diagnostics). Strips were positioned directly over the blood samples’ surface inside the wells after careful homogenization, and measurements were conducted immediately before and after data acquisition.

2.3.

Optical Coherence Tomography Measurements

For each blood sample, 100 consecutive acquisitions (B-scans) were performed at the same location with a commercial Spectral Radar OCT system OCP930SR (Thorlabs Inc.) with a central wavelength of 930 nm, spectral bandwidth (FWHM) of 100 nm, and spatial resolution of $6\mu \mathrm{m}$ (lateral and axial) in the air. Each B-scan was $2000\times 512\text{pixels}$ in size, covering 4 mm of the sample (laterally). Those B-scans were then analyzed via software for calculation of the attenuation coefficient. In order to correct the depth probed in the acquisitions, aliquots of $0.5\mu \mathrm{L}$ were collected from every sample and placed on glass slides for determination of the refractive index (RI).

Custom software was developed for the analysis using a simple exponential decay model for the attenuation of the OCT system based on Beer–Lambert’s law,²⁶ as follows:

Such an approach, albeit simplistic, is able to differentiate the effect of varying samples on the attenuation of the OCT signal. More robust methods have been described in the literature, and are able to account for system-dependent and multiple-scattering effects, which affect OCT sampling of turbid media.²⁷ However, as in this viability study, the samples were under a controlled environment and care was taken to avoid time- and temperature-dependent aberrations; the fundamental difference in the aliquots analyzed was the glucose concentration, our parameter of interest. As such, it is the major source of variance in the combined effect of whole blood on OCT signal attenuation. Therefore, the main goal in using the Beer–Lambert’s law approach was to perform fast and easy-to-compute analysis of attenuation that would enable the visualization of the overall trend of $\mu$ across the samples.

Once a B-scan is loaded in the software, the user selects a region of interest and its A-scans are first aligned through a peak-finding function to compensate for irregularities or displacements of the sample, then averaged. A fitting of the model in Eq. (1) is performed in the resulting A-scan, and parameters $I_0$, $\mu$, and C are obtained. For the present study, the value of interest is $\mu$. The process is repeated for all 100 B-scans of every sample.

2.4.

Speckles in Optical Coherence Tomography

Speckles in the OCT arise as a consequence of the sensitivity to the phase of the cross-correlation of the probed and reference optical fields.²⁸ Nevertheless, the speckle pattern has intensity fluctuations over time, dependent on the behavior of the scatterers inside the sample. This makes it possible to differentiate samples through an analysis of the OCT signal’s intensity.

One approach to such an analysis is the autocorrelation of the OCT signal. An area of the sample is probed, acquiring consecutive A-scans. A single point p in the A-scans is studied through time. Its intensity variations will alter the autocorrelation values, which will decay to zero (i.e., no longer correlated) at different rates for different behaviors of the speckle pattern.

For $N$ A-scans acquired consecutively, the normalized autocorrelation for the point p may be written as

This calculation is repeated for 20 different sets of 1024 consecutive A-scans, and the autocorrelation values are averaged. A new point $p$ is assigned, and another iteration is performed. The range of points evaluated is chosen via user input in custom software, and it determines different depths for analysis. The acquisitions were performed with the beam directed at a small angle ($\sim 8\mathrm{deg}$) from the normal to the blood surface, to eliminate specular reflections. All the samples were at room temperature.

As fluctuations of intensity occur rapidly in time, a custom OCT system (1325 nm central wavelength) was built, as described elsewhere,²⁹ achieving an 8-kHz acquisition rate. Also, as demonstrated in our previous work, an observation parameter ${\tau }_c$ is defined as the time interval it takes for the autocorrelation values to decay to $1/e$.²⁹ This approach is fast, and the value of ${\tau }_c$ is easily obtained and is still sensitive to changes in the temporal behavior of speckles, which makes it interesting for future in vivo applications.

This OCT system, however, was not used for total attenuation calculations, as the software developed for its signal acquisition was optimized for temporal analysis, with a lower axial resolution than the OCP930SR. Therefore, all attenuation studies were carried out with a 930-nm OCT.

2.5.

Statistical Analysis

Glucose measurements are arithmetic means with their respective standard deviations. The average attenuation coefficient is obtained from 100 measurements for every sample and presented as a function of glucose concentration. Attenuation coefficient data were submitted to the Kolmogorov–Smirnov test, which attested they were not to be drawn from a normally distributed population, so multiple mean comparisons were performed using the Kruskal–Wallis test with Dunn’s as posttest. Diluted blood was further submitted to $K$-means cluster analysis to identify groups of similar data.

Speckle decorrelation data are withdrawn from a normal distribution according to the Shapiro–Wilk normality test. Therefore, one-way analysis of variance was performed, followed by Bonferroni’s for a multiple comparison test.

All data analyses were performed with GraphPad Prism 5 and OriginPro 9.1 software, and we assumed statistically significant values for $p<0.05$.

3.1.

Attenuation Coefficient

Our results for the obtained glucose concentrations in whole blood for attenuation coefficient analysis are presented in Table 1. The resulting glucose concentrations deviated from our expectations, presumably due to precision limitations of instruments used during sample preparation (dilution and weighing).

Table 1

Whole blood glucose concentrations for each group (n=12). Values are presented as means±SD.

This first range of blood glucose concentrations displayed good correlation between the computed $\mu$ value and the blood sugar levels, as reported in Fig. 1.

Fig. 1

Attenuation coefficients calculated in relation to average glucose concentrations from whole blood after RI correction ($\mathrm{IR}=1.26$). Multiple comparison analysis showed $p<0.001$ between every group (Kruskal–Wallis with Dunn’s as post hoc). Red line represents the best linear correlation obtained, with slope of $3.55\mathrm{e}\text{-}5\mathrm{mg}/\mathrm{dL}·\mu \mathrm{m}$ with an adjusted $R^2$ of 0.98, which indicates that a linear model is suitable for our data. SE = standard error.

Our results show a trend of increase of the signal attenuation with respect to growing glucose levels. Nevertheless, studies in the literature report the use of glucose as an optical clearing agent in biological applications,^30,31 and OCT studies have shown a decrease in attenuation with increasing glucose concentrations.³² This is a consequence of RI matching between the plasma and the erythrocytes,³³ and is also due to aggregations and changes in size of red blood cells (RBCs), as was suggested by Tuchin et al.³⁴ However, it was also reported by Tuchin that the total attenuation coefficient of blood increased when mixed with glucose in vitro,³⁴ and related work shows the increase trend up to $\mathrm{10,000}\mathrm{mg}/\mathrm{dL}$³⁵ in vitro. Such a value, nonetheless, lies far outside the normal glucose levels recommended for euglycemic subjects (between 70 and $100\mathrm{mg}/\mathrm{dL}$ according to Standards of Medical Care in Diabetes³⁶). These apparently contrasting results, however, have been observed and explained by Bednov et al.³⁷ in the near-infrared region. The increase in $\mu$ with the addition of glucose is a direct result of the osmotic shock of RBCs when in the hyperglicemic condition, deforming their normal shape and directly affecting the scattering properties of blood.³⁸ Such a change in the scattering coefficient greatly surpasses the effects of index matching, raising the total attenuation of the signal, and affects both sedimented and nonsedimented blood. Furthermore, the aforementioned study shows that a higher glucose concentration means a higher osmotic shock effect, which linearly scales up attenuation, matching our observations. Nevertheless, the reported glucose levels tested exceed the physiologic range. Our results, thus, show that the $\mu$ increase trend is also true for lower glucose levels closer to normal glucose concentrations. Moreover, it also demonstrates the sensitivity of OCT to the osmotic shock of RBCs. Still, the linear trend observed due to osmotic effects, associated with our glucose levels (all within physiologic scope), explains the absence of the so-called concentration-dependent scattering,²⁷ which induces a nonlinear relation between scattering and concentration of scatterers in the medium.

However, it has been suggested that time after mixing plays a major role in the optical clearing effect in vivo,³¹ and that was also confirmed for osmotic shock,³⁷ which is temporary as the RBCs adapt to the hyperglicemic conditions after $\sim 10\min$. All our measurements were performed within 5 min after the addition of glucose, so all were affected by the discussed condition. After the period of adaptation of RBCs, the scattering coefficients normalize, and index matching promotes the results observed in optical clearing studies. Such an observation on the morphology of erythrocytes, then, conciliates both types of results observed in the literature.

Statistical analysis reported significant differences for the attenuation coefficient among all groups with $p<0.001$ for blood glucose concentrations in the range of 230 to $517\mathrm{mg}/\mathrm{dL}$. Due to the fact that spacing (in glucose levels) between our groups varies greatly, it is not possible to define a differentiating resolution for the approach, but differences as low as $65\mathrm{mg}/\mathrm{dL}$ (between groups 1 and 2) were detectable.

Despite our samples’ concentrations still surpassing the values of normoglycemic subjects, patients with diabetic ketoacidosis were reported to have plasma glucose levels above $250\mathrm{mg}/\mathrm{dL}$, and severe crises of hyperglycemia crises may reach values exceeding $600\mathrm{mg}/\mathrm{dL}$,³⁹ which places our tests within a biologically valid range.

The temporary characteristic of osmotic shock, allied with the fact that such an effect occurs only in vitro, is a limiting factor on the proposed technique and greatly hinders its clinical application. Notwithstanding this, the proposed methodology was still followed in the subsequent experiments to fully report on the observation of hyperglicemic shock through OCT.

In our second experiment with diluted blood, we have obtained more data points with different glucose concentrations. However, no linear correlation could be observed with this new set of data, as was the case for whole blood. The dilution of blood in isotonic saline solution to about 40% results in a stronger multiple scattering component in the OCT signal, as proposed by Popescu.²⁴ Associated with it is the fact that dilution directly affects the glucose concentration needed for the osmotic difference between RBCs and the serum to produce an appreciable change in the attenuation coefficient, which means that smaller additions of glucose will not be detectable by our approach, effectively lowering the resolution for differentiation. To confirm that hypothesis and search for correlations between $\mu$ and blood sugar, $K$-means clustering analysis was performed. The goal was to group together the concentrations with similar attenuation, enabling the observation of the necessary addition of glucose for differences noticeable by the technique. We were able to ascertain that there are three distinct groups in the range of glucose levels tested, as presented in Fig. 2.

Fig. 2

After RI depth correction ($\mathrm{IR}=1.45$), attenuation coefficients were calculated in relation to glucose concentrations from diluted blood and grouped in clusters by the K-means method.

If one considers the distance between group centers identified by the $K$-means, an approximate resolution may be defined as the average of those distances. In the reported case, the distances are $120\mathrm{mg}/\mathrm{dL}$ (cluster 1 to 2) and $146\mathrm{mg}/\mathrm{dL}$ (cluster 2 to 3), giving a resolution of $133\mathrm{mg}/\mathrm{dL}$ for the diluted samples, lower than the one obtained with whole blood. To ease visualization of those results, the average glucose concentration and $\mu$ of each group was calculated and plotted as new datapoints as a function of blood sugar levels. This plot is presented in Fig. 3, and the calculated values, along with standard deviation, are in Table 2.

Fig. 3

Mean attenuation coefficients calculated in relation to average glucose concentrations from diluted blood. Multiple comparison analysis showed $p<0.001$ between every cluster (Kruskal–Wallis with Dunn’s as post hoc). Red line represents the best linear correlation obtained, with an adjusted $R^2$ of 0.99 indicating our data can be represented by a linear model. SE = standard error.

Table 2

New groups resulting from our K-means analysis of diluted blood samples (n=12). Data are presented as mean±SD.

It is now possible to see once more the increasing tendency in the attenuation coefficient, and a linear fitting was performed. The fit reported a slope of 1.34e to $5\mathrm{mg}/\mathrm{dL}·\mu \mathrm{m}$ and an $R^2$ value of 0.99, suggesting that a linear model is adequate for our data. Furthermore, the clusters were subjected to the Kruskal–Wallis test with Dunn’s for multiple mean comparisons, and significant differences were observed between them ($p<0.05$).

3.2.

Speckle Decorrelation

The literature suggests that glucose levels directly affect blood viscosity.⁴⁰ Furthermore, viscosity of a medium is a determinant factor for the Brownian motion of particles suspended in such a medium, as indicated by the Stokes–Einstein equation for the diffusion coefficient. As the viscosity is inversely proportional to the latter, it is expected to be directly proportional to the decorrelation time of the Brownian motion. Autocorrelation analysis of the OCT signal has been shown to be able to discriminate between different blood viscosities due to salt¹⁹ and glucose in high concentrations.²⁰ The decorrelation times calculated by our method, therefore, are expected to be greater for the highest glucose concentrations.

The concentrations tested in this experiment are presented in Table 3. They were all aliquots from the same blood sample, and the reference glucose values were measured with the portable glucometer, which does not operate for values below $20\mathrm{mg}/\mathrm{dL}$. Group A did not receive glucose additions. The blood samples used were diluted with isotonic saline, as described previously, to introduce a stronger multiple scattering component to simulate conditions closer to the ones obtained when probing through tissue.

Table 3

Blood glucose concentrations for each group (n=15). Values are presented as means±SD.

Autocorrelation analysis was performed in regions $18\mu \mathrm{m}$ below the first peak of the signal, and each region covered $60\mu \mathrm{m}$ (10 points) in depth. There were intrasample variations in the decorrelation times calculated, which may be due to the rapidly decreasing signal for our system. Nevertheless, there is a clear trend of increase in the decorrelation values, as may be observed in the image plot performed on the matrix with the values, as shown in Fig. 4.

Fig. 4

Decorrelation times presented as a function of depth and blood glucose concentration (different groups) in false color mapping. Darker colors represent shorter times.

In Fig. 4, darker colors represent shorter decorrelation times, whereas brighter colors represent longer times. It is possible to see in that plot the tendency of the decorrelation times to be greater as the glucose levels get higher. Although each column does present variation in the color, they are still concentrated around certain tones, and such tones are lighter for the rightmost columns. This is a desired result for the technique and follows the expected outcome based on literature reports.

However, to observe only the difference intergroup, one may take the average of all the depths and plot the result as a representative value for each sample. This was done in Fig. 5.

Fig. 5

Average decorrelation times calculated for different blood glucose concentrations. One-way analysis of variance showed $p<0.05$ between every group (Bonferroni’s for multiple comparisons). Red line represents best linear fit.

Once more, the direct relationship between the decorrelation rate and glucose concentration is apparent. However, the data for $120\mathrm{mg}/\mathrm{dL}$ require attention. This point does not follow the trend observed for the remaining values and unfavorably affects the linear growth of decorrelation time through sugar levels. It is especially notable in the linear fit (red line) performed. Even with the error bars, the fit barely crosses this data point. Consequently, the $R^2$ for this fit is only 0.852, and the residual sum of squares is high. That may indicate that a linear model is not suitable for our data and may not be a good guess. However, as the remaining points do have a good linear fit ($R^2$ of 0.97), another possibility is the $120\mathrm{mg}/\mathrm{dL}$ sample may have been badly sampled, or each glucose concentration may need more analyzed points for the average to be a good representative. In such a case, the linear trend may be an acceptable model for the observed relation and cannot be disregarded. Further studies will be performed to check each hypothesis.

The obtained results suggest that high glucose levels increase the viscosity of the blood, which is in good agreement with other reports from the literature. The increase in the decorrelation time of speckles in the OCT has been observed before,²⁰ but the range of concentrations tested lies outside the range of normoglycemic subjects and commonly found glucose levels in diabetic patients.^20,36,39 We aimed to test the ability of OCT to differentiate blood sugar in lower concentrations, with the highest one studied in our trials ($\sim 355\mathrm{mg}/\mathrm{dL}$) coinciding with the lowest already reported ($\sim 360\mathrm{mg}/\mathrm{dL}$ or 20 mM). Nevertheless, the approach to the autocorrelation of our study also differs from the one already reported, being based on a parameter readily obtainable from the data, enabling a faster result but still able to discriminate the samples.

Our results thus, along with others works evaluating OCT’s signal autocorrelation capacity to detect changes in blood viscosity, indicate OCT as a potential tool for noninvasive diagnostics. Nonetheless, apart from the linear model validation, studies will also be conducted on the technique’s resolution for different glucose concentrations, which needs to be improved for a real application. Still, despite being performed ex vivo, those results are a base for a noninvasive in vivo approach for glucose monitoring with OCT. It has been demonstrated that the time-varying speckles originating from blood are still detectable by OCT in vivo,⁴¹ which further supports the possible application of speckle decorrelation analysis.

3.3.

General Considerations

Our results suggest that the attenuation coefficient is a feasible way to differentiate glucose levels in blood samples using OCT.

Nevertheless, the technique was studied only under osmotic shock conditions, and in vivo applications may not be viable, as these effects are only reproducible in vitro. Still, such results demonstrate the sensibility of OCT to the shape of RBCs and the consequent change in scattering it promotes on whole blood. Care must thus be taken when using OCT to study glucose-dependent parameters, as time after dilution plays a pivotal role in signal attenuation.

A greater variety of glucose concentrations may reveal the differentiating resolution of the method for whole blood, but with the data from the performed tests, differences of $65\mathrm{mg}/\mathrm{dL}$ were detected. However, in the presence of a greater multiple scattering component, as in the case of saline-diluted blood, this resolution was greatly impacted, and the average resolution was calculated to be $133\mathrm{mg}/\mathrm{dL}$, half the sensitivity reported for whole blood.

Also, the decorrelation time, defined as a simple parameter in the OCT autocorrelation signal, was shown to be sensitive to different glucose levels in the blood samples, as the addition of glucose alters the viscosity of the medium and therefore the Brownian motion of particles suspended in the plasma. Those distinct behaviors from the particles affect the speckle pattern observed in OCT and are enough to enable differentiation of blood sugar levels. However, in vivo studies are still necessary to check if the effects are still visible when under biological tissue. Notwithstanding this, the speckle decorrelation analysis here used is a novel approach for glucose monitoring and presents promising results, but is still at a preliminary stage. The reported findings are hopefully a starting point for further studies to be performed.

Finally, even though OCT systems are currently restricted to clinical and research environments, a trend to make those systems available for domestic use is observed in the literature.^42,43 Therefore, OCT may be used as a self-monitoring device by patients, and having a variety of analysis techniques already developed for this purpose is of great importance for the social impact of the technology.