1. Introduction
Non-invasive functional brain mapping is an imaging technique used to localize the functional areas of the patient brain. This technique is used during brain tumor resection surgery to indicate to the neurosurgeon the cortical tissues that should not be removed without cognitive impairment. Functional magnetic resonance imaging (fMRI) [
1] is the preoperative gold standard for the identification of the patient brain functional areas. However, after patient craniotomy, a brain shift invalidates the relevance of neuro-navigation to intraoperatively localize the functional areas of the patient brain [
2]. In order to prevent any localization error, intraoperative MRI has been suggested, but it complicates the surgery gesture, which makes it rarely used. For these reasons, electrical brain stimulation [
3] is preferred during neurosurgery. However, this technique suffers from limitations because the measurements could trigger epilepsy seizures. Since optical imaging combined with a quantitative modeling of brain hemodynamic biomarkers could evaluate in real time the functional areas during neurosurgery [
4,
5,
6], this technique could serve as a tool of choice to complement the electrical brain stimulation.
Hyperspectral imaging allows the in vivo monitoring of the hemodynamic and metabolic status of an exposed cortex. Hyperspectral imaging provides spatially and spectrally resolved images using numerous and contiguous spectral bands [
7]. In comparison, a standard color camera (or RGB camera) acquires three colors (red, green, and blue) using broad and overlapping spectral detectors. Both techniques have the ability to measure the oxygenation changes in the tissue using the modified Beer–Lambert law [
5,
8,
9,
10,
11,
12]. In functional brain mapping studies, the concentration changes of oxy- (
) and deoxy-hemoglobin (
) can be analyzed to identify the activated cortical areas [
5,
13,
14,
15,
16,
17,
18,
19,
20,
21]. The acquisition of the intrinsic signal in the near-infrared range offers the potential to monitor the brain metabolism with the quantification of the concentration changes of the oxidative state of cytochrome-c-oxidase (
) [
22,
23,
24,
25]. Hyperspectral and color cameras combined with a white light illumination are simple and powerful tools for the computation of intraoperative functional brain maps. The objective is to guide the neurosurgeon during brain surgery to prevent any functional impairments after surgical procedures (tumor resection).
In the literature, all wavelength bands acquired by hyperspectral imaging setups are used to measure the hemodynamic and metabolic changes in the brain [
10,
24,
25,
26]. However, there are some studies in which the choice of the selected spectral bands is discussed. Bale et al. [
23] showed that tens to hundreds of spectral bands acquired with a broadband near-infrared spectroscopy setup (780 nm to 900 nm) can be used to measure the oxCCO concentration changes. Arifler et al. [
27] showed that eight wavelength combinations between 780 nm and 900 nm give rise to the least possible estimation errors for the deconvolution of
,
, and
when compared to a gold standard (121 wavelengths included between 780 nm and 900 nm). Giannoni et al. [
28] proposed a Monte Carlo framework to investigate the performances of broadband spectroscopy to quantify the brain hemodynamic and metabolic responses. The results of this study indicated that eight wavelength between 780 nm and 900 nm should be selected to provide minimal differences in quantification compared to a gold standard of 121 wavelengths (780 nm to 900 nm). Sudakou et al. [
29] proposed a method based on the error propagation analysis and Monte Carlo simulations (three layer model: scalp, skull, and brain) allowing the estimation of the cytochrome-c-oxidase uncertainty in data measured with a multispectral time-resolved near-infrared spectroscopy device. The results of this study indicated that 16 wavelengths between 688 and 875 nm could be used to minimize the standard deviation of the cytochrome-c-oxidase concentration changes in the brain layer. Wavelength optimization problems have also been studied for other optical imaging techniques such as near-infrared optical tomography. Chen et al. [
30] identified seven laser diodes among 38 commercially available diodes in the range of 633–980 nm to estimate four chromophores (HbO
, Hb, water, and lipids) and the scattering prefactor in breast tissue. Chen et al. used the condition number and the residual norm to identify the optimal matrix used for the resolution of a linear system and thus to estimate four chromophores and the scattering prefactor in breast tissue. The optimal wavelengths were identified for a large residual norm and small condition number. The residual norm and the condition number can be interpreted as parameters representing the uniqueness and stability of the solution, respectively.
The commercial hyperspectral cameras have limited choices in the available spectral bands and do not have a spectral resolution as high as the broadband spectroscopy devices used by Bale et al. and Arifler et al. Therefore, the optimal spectral bands identified in these studies may not be available with the commercial cameras. Moreover, the more spectral bands are used, the more time is needed to compute functional brain maps. Since time is the key factor in intraoperative imaging, the smallest number of spectral bands must be acquired while ensuring minimal quantification errors.
First, we propose in this study a method to define the optimal spectral combinations of a commercial hyperspectral camera for intraoperative hemodynamic and metabolic brain mapping. This method could be used with any hyperspectral or standard RGB camera to evaluate its ability to compute accurate hemodynamic (
and
) and/or metabolic (
) brain maps following neuronal activation. The method is based on the Monte Carlo simulations of the acquisition of the intrinsic signal acquired by a camera. All spectral combinations of the hyperspectral camera are tested to evaluate the optimal spectral configuration that minimize the quantification errors in
,
, and
. In this work, we also show that a spectral correction [
10] of the reflection spectra acquired by a mosaic hyperspectral sensor is mandatory to minimize the chromophores’ quantification errors. Finally, we compare standard RGB imaging and hyperspectral imaging for hemodynamic and metabolic brain mapping. We demonstrate that RGB imaging is a low cost, but not an accurate solution to identify the functional areas in a patient brain based on the analysis of the cortical hemodynamics. Hyperspectral imaging is the ideal solution for an accurate computation of hemodynamic and metabolic brain maps.
3. Results
3.1. Determination of the Optimal Spectral Configuration of the Hyperspectral Camera
In
Figure 6, the quantification errors in
,
and
(see Equation (
10)) obtained for the best simulation-based deconvolution systems of
N spectral bands (
) are represented. The solid lines stand for the mean quantification error. The colored areas represent the dispersion range of the quantification errors. The vertical dashed lines indicate the optimal number of spectral bands for the deconvolution of each chromophore. Both corrected (in green) and uncorrected spectral configuration (in red) are represented. Note that these data were obtained using noisy measurements; zeros mean Gaussian noises were added to the camera intensities and to the Monte Carlo quantities (mean path length and reflection spectra); see
Section 2.6.1.
For the three chromophores, the quantification errors are important when a small number of spectral bands are used, but the quantification errors decrease with a larger number of bands. The quantification errors in
are on average
lower when the spectral correction is applied than when it is not applied. The quantification errors in
are on average
lower when the spectral correction is applied than when it is not applied. The quantification errors in
are equivalent with and without spectral correction. We can notice that the quantification errors reach a plateau when 15 corrected and uncorrected spectral bands are used. These spectral bands are indicated by violet points in
Figure 7 and will be named optimal reduced spectral bands in the rest of the paper.
For each chromophore, different optimal spectral combinations are found. For the quantification of
, the best spectral configuration of our hyperspectral camera is composed of 25 spectral bands if the spectral correction is applied and of 16 spectral bands if the spectral correction is not applied. These spectral bands are indicated by black vertical lines in the HbO
molar extinction spectra in
Figure 7. For these configurations, the quantification error (with Monte Carlo and intensity noise addition) is equal to
(mean ± standard deviation of the
measurements) when the spectral correction is applied and is equal to
when the spectral correction is not applied.
For the quantification of
, the best spectral configuration of our hyperspectral camera is composed of 24 spectral bands if the spectral correction is applied and of 21 spectral bands if the spectral correction is not applied. These spectral bands are indicated by black vertical lines in the Hb molar extinction spectra in
Figure 7. For these configurations, the quantification error (with Monte Carlo and intensity noise addition) is equal to
(mean ± standard deviation of the
measurements) when the spectral correction is applied and is equal to
when the spectral correction is not applied.
For the quantification of
, the best spectral configuration of our hyperspectral camera is composed of 21 spectral bands if the spectral correction is applied and of 25 spectral bands if the spectral correction is not applied. These spectral bands are indicated by black vertical lines in the oxCCO molar extinction spectra in
Figure 7. For these configurations, the quantification error (with Monte Carlo and intensity noise addition) is equal to
(mean ± standard deviation of the
measurements) when the spectral correction is applied and is equal to
when the spectral correction is not applied.
3.2. Impact of the Signal-to-Noise Ratio in the Measurements
As mentioned in
Section 2.3, the SNR values of the imaging system directly impact the amount of noise and the accuracy of the simulated quantities. To illustrate the effect of SNR on the measurements, the mean and standard deviation of the quantification errors in
,
and
(see Equation (
10)) obtained for the optimal spectral configuration of the hyperspectral camera (see
Figure 7) and the RGB camera are represented as a function of SNR; see
Figure 8 and
Figure 9. For each chromophore system, the mean and standard deviation values of the quantification errors are high for low SNR values and decrease with increasing SNR values.
When a two-chromophore system is considered with the RGB camera (SNR = 10), the quantification errors in and are equal to and , respectively. When the SNR value is equal to 1000, the quantification errors in and are equal to and , respectively. For high SNR values, it is interesting to note that the quantification of is not accurate (low values and high values) and the quantification of is accurate. Moreover, the mean quantification errors in reach a plateau for an SNR value equal to 110. For the hyperspectral camera with an SNR value equal to 10, the quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and when it is not applied. When the SNR value is equal to 1000, the quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and when it is not applied. For high SNR values, the quantification of and is accurate when the spectral correction is applied and not accurate when the spectral correction is not applied (low values and high values). When the spectral correction is not applied, the mean quantification errors in and reach a plateau for an SNR value equal to 200 and 110, respectively.
When a three-chromophore system is considered with the RGB camera (SNR = 10), the quantification errors in , , and are equal to , , and , respectively. When the SNR value is equal to 1000, the quantification errors in , , and are equal to , , and , respectively. The mean quantification errors in , , and reach a plateau for an SNR value equal to 110, 130, and 130, respectively. For high SNR values, the quantification of , , and is not accurate (low values and high values). For the hyperspectral camera with an SNR value equal to 10, the quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and when it is not applied. When the SNR value is equal to 1000, the quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and when it is not applied. The quantification errors in are equal to when the spectral correction is applied and 91 % when it is not applied. For high SNR values, the quantification of and is accurate when the spectral correction is applied and not accurate (low values and high values) when the spectral correction is not applied. The quantification of is not accurate with and without spectral correction (low values and high values). We can notice that the measurement uncertainty of is similar for all SNR values with and without spectral correction. When the spectral correction is not applied, the mean quantification errors in and reach a plateau for an SNR value equal to 200.
3.3. Hemodynamic Monitoring
The hemodynamic monitoring (
and
measurements) following a simulated patient neuronal stimulation (see
Section 2.7) is represented In
Figure 10 and
Figure 11. In
Figure 10, the
and
values were computed with the optimal spectral combination of the hyperspectral camera; see
Figure 7. In
Figure 11, the
and
values were computed with the RGB camera. In these figures, the modified Beer–Lambert law was computed
times, using different Monte Carlo and intensity noise occurrences for each time iteration. When using the 121 wavelengths included between 780 nm and 900 nm [
23] to quantify the concentration changes during the stimulation period (from
s to
s), there is a
overestimation of the
values and a
underestimation of the
values compared to the theoretical measurements. When using the eight wavelengths identified by Arifler et al. [
27] to quantify the concentration changes during the stimulation period, there is a
overestimation of the
values and a
underestimation of the
values compared to the theoretical measurements.
In
Figure 10, the hemodynamic monitoring following a simulated neuronal stimulation was computed using hyperspectral imaging. The optimal corrected and uncorrected spectral configurations of the hyperspectral camera were used to quantify the
and
values. The quantification dispersion ranges of the corrected and uncorrected spectral configurations have approximately the same range of magnitude. The standard deviation averaged over all time measurements for
and
are equal to
mol·L
and
mol·L
, respectively. When the spectral bands are corrected, the
and
values averaged over the
noisy measurements have a good match with the theoretical hemodynamic responses. During the stimulation period, there is on average a
overestimation of the
values and a
underestimation of the
values compared to the theoretical measurements. When the spectral bands are not corrected, there is on average a
underestimation of the
values and a
underestimation of the
values compared to the theoretical measurements. The correlation coefficients computed between the theoretical HbO
response and the
time courses are higher when the spectral bands are corrected (
) than when they are not (
). In the same way, the correlation coefficients computed between the theoretical Hb response and the
time courses are higher when the spectral bands are corrected (
) than when they are not (
). The correlation coefficient values are summarized in
Table 4.
In
Figure 11, the hemodynamic monitoring following a simulated neuronal stimulation was computed using RGB imaging. The standard deviation averaged over all time measurements is equal to
mol·L
and
mol·L
for
and
, respectively. The
values values averaged over the
noisy measurements do not have a good match with the theoretical HbO
response; however, the match is rather good between the
values and the theoretical Hb response. During the stimulation period, there is on average a
underestimation of the
values and a
underestimation of the
values compared to the theoretical measurements. The correlation coefficients computed between the theoretical HbO
response and the
time courses are equal to
. The correlation coefficients computed between the theoretical Hb response and the
time courses are equal to
. The correlation coefficient values are summarized in
Table 4.
3.4. Hemodynamic and Metabolic Monitoring
The hemodynamic and metabolic monitoring (
,
, and
measurements) following a simulated patient neuronal stimulation (see
Section 2.7) are represented in
Figure 12 and
Figure 13. In
Figure 12, the
,
, and
values were computed with the optimal spectral combination of the hyperspectral camera; see
Figure 7. In
Figure 13, the
,
, and
values were computed with the RGB camera. In these figures, the modified Beer–Lambert law was computed
times, using different Monte Carlo and intensity noise occurrences for each time iteration. When using the 121 wavelengths included between 780 nm and 900 nm [
23] to quantify the concentration changes during the stimulation period (from
s to
s), there is a
underestimation of the
values, a
overestimation of the
values, and a
overestimation of the
values compared to the theoretical measurements. When using the eight wavelengths identified by Arifler et al. [
27] to quantify the concentration changes during the stimulation period, there is a
underestimation of the
values, a
overestimation of the
values, and a
overestimation of the
values compared to the theoretical measurements.
In
Figure 12, the hemodynamic and metabolic monitoring following a simulated neuronal stimulation was computed using hyperspectral imaging. The optimal corrected and uncorrected spectral configurations of the hyperspectral camera were used to quantify the
,
, and
values. The quantification dispersion ranges of the corrected and uncorrected spectral configurations have approximately the same range of magnitude. The standard deviation averaged over all time measurements for
,
, and
are equal to
mol·L
,
mol·L
, and
mol·L
, respectively. When the spectral bands are corrected, the
and
values averaged over the
noisy measurements have a good match with the theoretical hemodynamic responses. The
values averaged over the
noisy measurements have a good match with the theoretical metabolic response with and without spectral correction. When the spectral bands are corrected, there is on average a
underestimation of the
values, a
overestimation of the
values, and a
overestimation of the
values compared to the theoretical measurements. When the spectral bands are not corrected, there is on average a
underestimation of the
values, a
underestimation of the
values, and a
underestimation of the
values compared to the theoretical measurements. The correlation coefficients computed between the theoretical HbO
response and the
time courses are higher when the spectral bands are corrected (
) than when they are not (
). The correlation coefficients computed between the theoretical Hb response and the
time courses are higher when the spectral bands are corrected (
) than when they are not (
). The correlation coefficients computed between the theoretical oxCCO response and the
time courses are higher when the spectral bands are corrected (
) than when they are not (
). The correlation coefficient values are summarized in
Table 5.
In
Figure 13, the hemodynamic and metabolic monitoring following a simulated neuronal stimulation was computed using RGB imaging. The standard deviation averaged over all time measurements is equal to
mol·L
,
mol·L
, and
mol·L
for
,
, and
, respectively. The computed
,
, and
values do not have a good match with the theoretical hemodynamic and metabolic responses. During the stimulation period, there is on average a
underestimation of the
values, a
underestimation of the
values, and a
overestimation of the
values compared to the theoretical measurements. We can notice that there is an important crosstalk between the chromophores. The
values were mainly interpreted as Hb variations, the
values as oxCCO variations, and the
as HbO
variations.
4. Discussion
Matcher et al. [
46] showed that the performance of spectroscopic analysis can be improved by increasing the number of wavelengths of illumination. The results of our study are consistent with Matcher et al.’s results; see
Figure 6. The simulation-based method presented in this paper aimed to identify the optimal spectral combination of our hyperspectral camera for the brain hemodynamic and metabolic monitoring. When the spectral bands are corrected, the optimal spectral configuration is composed of 21 and 22 spectral bands; see
Figure 7. This configuration could be however reduced from 10 to 12 spectral bands while keeping fairly constant performances. We can observe a plateau in the quantification errors of
when 10 or more spectral bands are used. We also can observe a plateau in the quantification errors in
and
when 12 or more spectral bands are used; see
Figure 6. This reduction of the number of the spectral bands could be interesting for the real-time computation of hemodynamic and metabolic brain maps in the operative room using commercial hyperspectral cameras and for the conception of dedicated cameras used in functional brain mapping studies.
In this study, the optimal spectral configurations were obtained by searching the three chromophores deconvolution systems that minimize the quantification errors in
,
, and
. Therefore, three different spectral configurations were obtained for the deconvolution of the three chromophores. These spectral configurations, as well as the reference configuration identified by Bale et al. [
23] and Arifler et al [
27] are represented in
Figure 14.
The comparison between the spectral configuration identified by Arifler and al. and ours is not trivial since our hyperspectral camera does not acquire exactly the same wavelengths; see
Figure 2. There are some similarities between the wavelengths used in our system and in the system of Arifler et al. Indeed, between 780 nm and 900 nm, some wavelengths correspond. Our system also used others wavelengths inferior to 780 nm (703, 754, and 767 nm) and superior to 900 nm (916, 929, 937, 947, 953, and 958 nm). The quantification of
,
, and
computed with the optimal spectral bands of our hyperspectral camera are consistent with those computed with 121 wavelengths included between 780 nm and 900 nm [
23] and with those computed with the eight wavelengths identified by Arifler et al. [
27]. For a two-chromophore deconvolution system, the quantification error in
measured with our system is equivalent to those measured with 121 wavelengths. However, the choice of our spectral bands aims to reduce the quantification error in
; see
Figure 10. For a three-chromophore deconvolution system, the quantification errors measured with our system are a little higher than those measured with 121 wavelengths or with Arifler et al.’s wavelengths; see
Figure 12. This difference may be explained because the illumination and the acquisition of the 121 wavelengths of reference, as well as the eight wavelengths identified by Arifler et al. were simulated as ideal sources and detectors. This means that for these simulations, the term
in Equation (
7) is equal to one. To efficiently compare our spectral configuration with these spectral configurations of reference, the spectral sensitivities of the detectors and the illumination sources have to be considered. However, the acquisition condition is different since we acquired the intrinsic signal of an exposed cortex, whereas the wavelengths identified in Bale et al.’s and Arfiler et al.’s studies were identified for functional near-infrared spectroscopy devices.
We showed that a spectral correction (see
Section 2.2) is required when a hyperspectral mosaic sensor is used [
10]. The spectral correction aims to reduce the quantification errors in the measurements and allows a better follows-up of the temporal hemodynamic and metabolic variations; see
Table 5.
We also compared hyperspectral imaging to RGB imaging for the computation of hemodynamic and metabolic brain maps. Hyperspectral imaging is the suitable solution to compute hemodynamic maps and metabolic maps thanks to its ability to acquire the intrinsic signal in the near-infrared range [
23]. A very important crosstalk between HbO
, Hb, and oxCCO can be observed when the RGB camera is used for the computation of hemodynamic and metabolic brain maps; see
Figure 13. Therefore, RGB imaging is not a suitable solution to compute metabolic brain maps. An RGB camera could be however a low-cost solution to compute hemodynamic maps; see
Figure 11. This solution is not accurate to quantify
, but is very accurate for the quantification of
. This original result is interesting because most surgical microscopes used in the operating room are equipped with standard RGB cameras. It is known that the BOLD signal used in fMRI studies is predominately due to the paramagnetic properties of deoxygenated hemoglobin [
47]. This result indicates that the
quantified with RGB imaging can be used in a robust way for intraoperative functional mapping based on SPM analyses.
We incorporated the camera noises in our simulation to find the most robust and reliable spectral configuration. In
Figure 8 and
Figure 9, we show that the SNR of the imaging system has a drastic impact on quantification performances. An accurate quantification could be obtained with a high SNR value. Therefore, the light source and the camera specifications and settings have to be carefully chosen in order to guarantee an optimal SNR. This simulation framework could be a great tool for industry or researchers working on intraoperative functional brain mapping solutions to help them in the choice of commercial camera. However, our simulation framework needs to be improved. For the moment, a homogeneous volume of grey matter was considered. A realistic mapping of the exposed cortex could be simulated as suggested by Giannoni et al. [
28]. This will be considered in future studies. Moreover, the hemodynamic and metabolic changes following neuronal activation were homogeneously simulated in the volume of grey matter. These events are obviously not consistent with those appearing in a real cortical tissue. This modeling also has to be taken into account to improve our method of identification of the optimal spectral bands of a hyperspectral camera for brain hemodynamic and metabolic monitoring.