Combined autofluorescence and diffuse reflectance for brain tumour surgical guidance: initial ex vivo study results.

This ex vivo study was conducted to assess the potential of using a fibre optic probe system based on autofluorescence and diffuse reflectance for tissue differentiation in the brain. A total of 180 optical measurements were acquired from 28 brain specimens (five patients) with eight excitation and emission wavelengths spanning from 300 to 700 nm. Partial least square-linear discriminant analysis (PLS-LDA) was used for tissue discrimination. Leave-one-out cross validation (LOOCV) was then used to evaluate the performance of the classification model. Grey matter was differentiated from tumour tissue with sensitivity of 89.3% and specificity of 92.5%. The variable importance in projection (VIP) derived from the PLS regression was applied to wavelengths selection, and identified the biochemical sources of the detected signals. The initial results of the study were promising and point the way towards a cost-effective, miniaturized hand-held probe for real time and label-free surgical guidance.


Introduction
Glioblastoma account for approximately 80% of all primary malignant brain tumours [1]. They are a particularly infiltrative type of tumours and their treatment involves surgical resection followed by chemotherapy and radiotherapy [2][3][4]. The efficiency of resection is a primary indicator of post-operative outcomes. Unfortunately, the infiltrative nature of glioblastoma makes the intraoperative discrimination between diseased and healthy brain tissue challenging. As a result, complete resection cannot be guaranteed. At present, tumour localization is achieved through preoperative or intraoperative based on Magnetic Resonance Imaging (MRI), Computed Tomography (CT) or neuro-navigation. Although effective, these methods have limited degree of accuracy due to shifting of the tissue once the cranial pressure is released as well as extending duration of the procedure. As such, there is much need for a quantitative, real-time method to differentiate tumour tissue from healthy brain.
Optical techniques have been widely investigated for neuro-oncological applications [5][6][7][8][9][10] and fluorescence spectroscopy has been shown to have potential for intraoperative delineation of tumour resection margins [11][12][13][14][15][16]. Fluorescence spectroscopy can detect endogenous fluorophores [17,18] or preoperatively administered contrast agents such as 5-aminolevulonic acid (5-ALA), the latter is widely used to delineate tumour tissue in clinical practice [19][20][21][22][23]. 5-ALA given by oral route is taken up by glioma cells through the disrupted blood-brain barrier and metabolized into a fluorescent protoporphyrin IX (PpIX) molecule [24]. In the diseased cells, the PpIX accumulated at relatively high concentration and can be visualized under intraoperative fluorescence microscope [25][26][27]. 5-ALA assisted surgery has been shown to be very successful but it has some limitations [21,28,29]. For example, PpIX fluorescence can be observed in the area of peritumoral edema or inflammatory cells and reactive astrocyte infiltration with or without tumour cells, which increases the risk of excessive resection [29]. In addition, photobleaching of PpIX fluorescence during neurosurgery can make detection of diffuse tumour margins subjective and challenging [30]. To avoid some of the shortcomings of 5-ALA assisted surgery, alternative, label free, autofluorescence spectroscopy methods have been investigated [31][32][33]. Such techniques are designed to detect endogenous, tissue specific, fluorescent biomarkers such as reduced nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), lipopigments, and porphyrin [5]. Furthermore, by combining autofluorescence with diffuse reflectance spectroscopy (DRS) several groups have attempted to differentiate tumour from healthy breast [34][35][36], liver [37,38] and brain tissue [39,40]. For example, Lin et al. [39] used a 337 nm nitrogen laser to induce biomarker fluorescence at 460 nm in an in vitro brain tumour study. In this work they collected reflectance data between 400-850 nm and used it to account for tissue optical properties, this enabled them to differentiate grey matter and brain tumours with average sensitivity of 96% and specificity of 93%. In another study, Xie et al. [41] developed a fibre optic instrument with fluorescence and DRS capability for photodynamic therapy. The optical measurements in this study were processed using non-linear least squares support vector machines resulting in diagnostic precision of 100% for in vivo skin tumour with topically applied 5-ALA.
Previous investigations have primarily concentrated on fluorophore detection using a limited number of excitation-emission bands. Unlike high resolution spectrometers, the narrow bandwidth allowed these systems to perform measurements in fractions of seconds, making them less sensitive to background illumination and more suitable to clinical environments. However, the limited optical bandwidths limit the number of fluorophores that can be detected and makes it difficult to account for the specific tissue optical properties that may influence the biomarker excitation and fluorescence intensity. The use of multiple excitation and detection bands have the potential to improve the accuracy of intrinsic tissue fluorescence measurements by more accurately compensating for tissue absorption and scattering. Such approach could potentially improve the predicative algorithm for tissue recognition to achieve positive surgical outcomes.
Hence, in this study, an autofluorescence and DRS-based system for tissue differentiation in the brain was assessed. Using eight excitation and emission wavelengths spanning from 300 to 700 nm, a total of 180 optical measurements were acquired from 28 brain samples (five patients). Partial least square-linear discriminant analysis (PLS-LDA) was used as the classification algorithm. Leave-one-out cross validation (LOOCV) was then used to evaluate the performance of the classification model. The variable importance in projection (VIP) derived from the PLS regression was applied to source/detector wavelength pairs selection. The system described in the following sections is a useful step towards a miniaturized, hand-held integrated device that uses multi-band spectroscopy for intraoperative identification of brain tumour tissue. Although the initial results reported here are promising, future clinical studies with larger patient population will be necessary to validate these observations and improve the discriminant algorithm.

Monte-Carlo simulation
To predict the behaviour of excitation light distribution in different types of brain tissue, Monte-Carlo simulations of light transport in turbid media were performed using the previously developed MCML code [42]. In each simulation, 10 7 photon packages were injected. The optical properties described by Yaroslavsky et al. [43] were used in the simulation and are summarized in Table 1. The two wavelengths selected for simulation are based on illumination wavelengths used in this study. Simulation at 405 nm aimed to provide a rough picture of light distribution in brain tissue between 300-530 nm, whilst simulation at 600 nm provides information for light distribution between 530-700 nm.

Brain tissue samples for ex vivo studies
A total of 28 glioma brain tissue samples were acquired from five patients undergoing tumour resection surgery. The biopsies were collected with informed consent and in accordance with guidance from the Clinical Research and Ethics Committee in University College Cork (UCC), Ireland. The median age was 48 years (ranging from 34-76 years) with two females and three males. The corresponding pathology appearance of each optical measurement is summarized in Table 2. The tumour was graded in a qualitative manner as being of high, moderate or low cellularity.

Instrumentation
The instrumentation used in this study is schematically illustrated in Fig. 1(a). Briefly, the system incorporated a light source with fibre-coupled light emitting diodes (LEDs), a fibre optic probe bundle, an optical components and detection unit, a data acquisition board, and a computer. The light source comprised of eight fibre-coupled LEDs (Thorlabs, Germany) at wavelengths: 300, 340, 395, 470, 505, 530, 595 and 700 nm. Each of the LEDs was driven using a dedicated, single channel driver (LEDD1B, Thorlabs, Germany) and controlled using LabVIEW. All LEDs worked were pulsed in sequence of 3 ms. To allow ambient background measurements, each illumination sequence included two additional 3 ms intervals, at the start and at the end where all LEDs were off. Single illumination sequence lasting 30 ms was repeated 100 times resulting in total measurement period of 3 seconds. The customized multi-fibre probe (FibreTech, Optica, Canada) consisted of eight illumination fibres surrounding a single light collecting fibre, bundled together into a stainless-steel tube with diameter of 10 mm ( Fig. 1(a)). The core diameters of the collecting fibre and illumination fibres were 600 µm and 400 µm, respectively. Numerical aperture (NA) in both cases was 0.22. The centre-to-centre distance between the light illumination and collection fibres was 0.56 mm.
The collected light, carrying the DRS and fluorescence signals, was patched to a detector array where it was collimated and then split into eight different light paths using seven dichroic beam splitters (Semrock, Laser 2000, Cambridgeshire, UK). This method of wavelengths separation ensured that the DRS signal with shortest wavelengths, matching that of illumination, was patched to a single APD. The remaining red-shifted fluorescence signals travelled further through the dichroic, reflected one by one into corresponding APDs. The light reflected from each dichroic surface, was filtered using an optical bandpass filter (300, 340, 395, 470, 505, 530 and 700 nm, Semrock, Laser 2000, Cambridgeshire, UK) to attenuate crosstalk between channels and then focused using a condenser lens on an APD (APD4102A-M, Thorlabs, Germany). Figure 1(b) shows an example of optical signal collected by the avalanche photodiode detectors (APDs) in one cycle.
The analog output from each APD was selectively amplified using a custom, 8-channel amplifier (Grenmore, Ireland). The gain on channels where an APD was measuring diffuse reflectance was set to x1 and the gain on the channels where an APD was measuring autofluorescence was set to x100. Following amplification, the signals were sampled at 30 kHz and digitized using an NI-9205 module.

Optical measurements
Prior to every set of measurements on a tissue sample, a set of standard measurements were performed on a set of reflectance and fluorescence standards (Spectralon, Labsphere Inc, North Sutton, US). The tissue sample was then placed onto the sample stage, positioned under the camera crosshairs using the stage X-Y adjustment, and a digital image was recorded. Next, the entire stage was moved under the fibre-optic probe which was indexed to the crosshairs of the camera and made it possible to record the location of optical measurements on the tissue sample. The tip of the fibre probe was brought into gentle contact with the target tissue ( Fig. 2(a)). If tissue sample size allowed, this procedure was repeated, and spectra were recorded from multiple sites. Following spectral acquisition, the samples were transferred into a tissue fixing cassette. It was ensured that the orientation of the sample in the cassette was the same as on the stage. This allowed the fixed samples to be sectioned in the same plane as digital images (Fig. 2(b)). A histopathological diagnosis was then acquired for each measurement site by cross-referencing the location of the crosshairs on the digital image to the fixed sample section. The tissue type identified by the neuropathologist at review was considered to be the ground truth.

Data pre-processing
Several steps were conducted to pre-process the collected pulses on all APDs before statistical analysis (Fig. 3). First, a baseline correction was performed with light source blocked by subtracting the DC offsets measured separately. To increase the signal-to-noise ratio, the data was integrated over 100 illumination cycles, corresponding to a total integration time of 3 seconds. Then, to remove artefacts associated with light source and amplifier switching, the first 20 data points and the last 5 data points associated with each light source interval were removed, and remaining 65 data points were integrated to a single average value. Thereafter, the average value of the ambient background was subtracted from all other signals.  3. Flowchart of data pre-processing that was retrieved to analyse brain tissue optical signal.
The intrinsic fluorescence signal compensated for the attenuation of the excitation light by forming the ratio of fluorescence emission and reflectance at excitation (Eq. (1)).
where F (λ emission ) is the "raw" fluorescence intensity excited by λ excitation ; R (λ excitation ) is the reflectance intensity at the excitation wavelength.
In order to correct for day-to-day variations in detector sensitivity and light source power, a correction factor C i (λ) was determined by taking the ratio of the reference spectrum intensity data vector for i-th patient (S (ref ,i) (λ)) to that of the first patient (S (ref ,1) (λ)) at each excitation/emission wavelength. That is, where λ represents each source/detector wavelength pair. Each correction factor, C i (λ), was then divided by each measurement intensity of data vector generated described in above section for i-th patient.
After the DC baseline, ambient light, excitation attenuation, and instrument variation corrections were applied, the data was collated and a single n × p matrix was generated. Each row (n) corresponded to optical measurement from one investigation site and each column (p) corresponded to a variable (source/detector wavelength pair of the system).

Statistical analysis
In this work, PLS-LDA was chosen for classification of brain tissue types. In this method, PLS was first performed over the whole data matrix, then a supervised classification scheme was developed using LDA model. Unlike PLS discriminant analysis (PLS-DA) which is based on PLS regression algorithm by using dummy levels for the classes and putting a threshold on the regression results, PLS-LDA uses LDA as a classification model constructed using score matrix generated by PLS regression as input. PLS-LDA uses the number of PLS latent variables to adjust the LDA projection direction, which helps to avoid overfitting. Due to the relatively small sample size, leave-one-out cross validation (LOOCV) was then used to evaluate the performance of the classification model built using the PLS-LDA method. The root mean squared errors of cross validation (RMSECV) was used to optimise the models.
The variable importance in projection (VIP) was used as variable selection method and applied in system parameter (source/detector wavelength) selection. Student's t-test at a 5% significance level was further performed on the reflectance and fluorescence signal intensity at selected source/detector wavelength pairs to see if any observed differences in these parameters between tissue types were statistically significant.
All data in this paper and numerical analysis were processed using MATLAB 2019b (The MathWorks Inc. Natick, Massachusetts). All statistical analysis was conducted using the Statistics and Machine Learning Toolbox function in MATLAB. Figure 4 shows the Monte-Carlo simulation of fluence rate distribution in grey matter, white matter and glioma at wavelength of 405 nm and 600 nm based on optical properties summarized in Table 1. There is a difference between the fluence rate distribution in grey matter and glioma at both wavelengths, which is promising for differentiating tissue types. As expected light propagation into tissue is attenuated by both tissue scattering and absorption and is wavelength dependent. The tissue volume in which fluorophores are excited is approximated as the volume in which the excitation light is distributed.

Monte-Carlo simulation
The simulated reflectance R(r) data as a function of radial position is presented in Fig. 5. It is notable that there is little variation in a pivot point region of grey and white matter curves on reflectance (0.6 mm at 405 mm, 1 mm at 600 nm). This indicates that the reflectance values predicted from these simulations of grey and white matter using the source-detector distance employed in our measurements of 0.56 mm are close, while the reflectance level of glioma is noticeably lower than that of grey and white matter. However, it should be noted that as Monte-Carlo simulation is a numerical analysis, the results are highly dependent on the optical parameters input. Optical properties for human brain tissue reported in the literature vary due to disparities between experimental method limitations or sample preparation details [43,44]. Therefore, verifying experimental work is essential.

DRS and autofluorescence signal profiles of optical measurements
A total of 180 optical measurements -53 on grey matter, 87 on white matter, 28 on tumour tissue and 12 on necrotic tissue -were collected from 28 brain specimens of five brain tumour patients. Figure 6 shows average optical signals of grey matter and tumour tissue after subtraction of ambient background, compensation for day-to-day system response variation ( Fig. 6(a)), and correction for excitation light attenuation ( Fig. 6(b)). The optical signals of white matter and necrotic tissue are excluded from the plot for readability and included in the Supplement 1 Figure  S1. Error bars represent the standard deviation of the means for each type of tissue. Due to the broad spectral bandwidth of LEDs at wavelength 470 nm, 505 nm and 530 nm, the adjacent detection wavelength channels showed a significant amount of crosstalk. Therefore, five data  The mean reflected intensities of grey matter show no significant difference with tumour at any wavelengths. Figure 6(b) shows fluorescence signal profile after correcting for the light attenuation. The intensity of the 470/700 nm fluorescence from tumour is less than that from grey matter. The tissue absorption and scattering properties of grey matter and tumour are quite different at low wavelengths (refer to Table 1). As mentioned above, tissue can attenuate the intrinsic fluorescence intensity by absorption and scattering, using reflectance-normalization to compensate helps to approach true autofluorescence. As can be seen, Fig. 6(b) correcting for the excitation attenuation changes the optical signal profiles. At the fluorescence region excited by 300 nm, the corrected signals from tumour show higher intensities than that from grey matter. The classification of tissue types described below is based on these corrected signals. Figure 7 shows the RMSECV plotted against the number of PLS components for classifying grey matter and tumour. Choosing the number of PLS components is a critical step in building a PLS-LDA model. The "elbow" of the plot was chosen as the optimal one [45]. It can be seen that, RMSECV does not change significantly after first three components. Hence, the first three PLS components were selected to build the classification model. Table 3 displays the confusion matrix of the classification results using leave-one-out cross validation. The diagonal values (highlighted in bold font) reflect the percentage of correct tissue type classification. Grey matter can be differentiated from tumour tissue with sensitivity of 89.3% and specificity of 92.5% (95% confidence intervals).

System parameters (source/detector wavelengths) selection
The variable importance in projection (VIP) derived from the PLS regression was used as variable selection method. The figures are calculated as the weighted sum of squares of the PLS weight W (weight being the explained variance of each latent variable). The VIP that results in classification in grey matter and tumour is shown in Fig. 8. Overall, the top ranked variables that contributed most for classifying grey matter and tumour are fluorescence at 470/700, 300/340, 300/395, 300/595, 395/530 nm and reflectance at 300, 395, 700 nm. Student's t-test was further performed on the reflectance and fluorescence signal intensity at these system parameters to see if any observed differences between grey matter and tumour were statistically significant. Figure 9 shows the boxplots of the reflectance and various autofluorescence signal at selected source/detector wavelength pairs in grey matter and tumour. Fig. 9. Boxplots of (a) DRS at 300, 395 and 700 nm, (b) fluorescence at different system parameters in grey matter and tumour tissue. (A significantly difference between tissue types indicated by "*" (p < 0.05) and "**" (p < Bonferroni corrected p-value as 0.0016), "ns" represents no statistically significant difference).
For reflectance measurements, no significant difference between both tissue types at 395 nm. At 300 nm and 700 nm, diffuse reflectance of grey matter was observed to be slightly (∼25%) greater than that of tumour tissue. No significant difference was observed between the tissue types for the fluorescence channels 300/595 and 395/530 nm. For the 470/700 nm channel, the intensity from tumour is 84% lower than that from grey matter, while at 300/340 and 300/395 nm, the intensity from tumour was 32% and 57% higher than that from grey matter, respectively. Generally, a p-value less than 0.05 is considered statistically significant. However, the risk of a type I error (false-positive) increases when making multiple statistical tests. To mitigate against this, a Bonferroni correction was conducted -dividing the original α-value (0.05) by the number of statistical analyses performed (31 variables). A Bonferroni corrected p-value calculated was 0.0016, any observed p-value less than 0.0016 was declared to be statistically significant. Under this criterion the 470/700 and 300/395 nm channels show a statistically significant different between grey matter and tumour.
Additional data are presented in Supplement 1 Figure S2, Figure S3 and Table S1 in Supplement 1.

Discussion and conclusion
In the present ex vivo study, autofluorescence and diffuse reflectance optical signals were acquired from 28 brain specimens (five patients) with eight excitation and emission wavelengths spanning from 300 to 700 nm. We present the classification results that grey matter could be distinguished from tumour tissue with sensitivity of 89.3% and specificity of 92.5% using leave-one-out cross validation approach with PLS-LDA algorithm.
In general, the scattering coefficient of white matter is higher than that of grey matter and brain tumour [43,46]. Hence, it is expected that the DRS signals from white matter is higher than that from tumour; this was what was observed in our study, white matter generally displayed 50% -95% higher reflectance intensity than that of tumour and grey matter at eight source wavelengths (Supplement 1, Figure S1). These results are in agreement with those ex vivo or in vivo studies which have been reported previously on human brain tissue [39,40]. The diffuse reflectance from grey matter showed no significant differences with tumour at most wavelengths ( Fig. 6). At wavelength of 700 nm where blood absorption has less influence, reflectance of grey matter was slightly higher, ∼25%, than that of tumour. These results are mostly in agreement with Monte-Carlo simulation demonstrated in Fig. 5(b). The intensity of the diffuse reflectance is correlated to the absorption and scattering coefficients of tissue samples. The same R(r) data may be obtained from two samples with different optical properties (e.g. a pivot region of white matter and grey matter curves shown in Fig. 5). Moreover, probe geometry, such as probe core dimension and source-detector distance, could also affect the irradiance of the light that reaches and collects from the tissue. With increasing source-detector distance, the detected light intensity declines exponentially (Fig. 5), but light propagate into tissue deeper thus the measurements to the brain tissue is more sensitive. Therefore, it should be balanced between sensitivity and signal strength in source-detector distance selection in visible and near infrared spectroscopy probe design. The centre-to-centre distance between the light illumination and detector fibres used in this study was 0.56 mm, which is a short distance and the collected light is predominantly reflected from tissue superficial layer.
Our experimental results confirm the finding from previous studies that DRS alone is not sufficient to discriminate grey matter from tumour tissue (Fig. 6). The VIP analysis (Fig. 8) suggested that the fluorescence signals at wavelengths 470/700 nm, 300/395 nm and 300/340 nm of grey matter and tumour were the most discriminative and the Bonferroni corrected p-value ( Fig. 9(b)) showed that two wavelength pairs 470/700 and 300/395 to be statistically significant. The fluorescence intensity at 470/700 nm from tumour was 80% lower than that from grey matter. This could correspond with porphyrin, a main endogenous fluorophore in the brain. Croce et al. [31] reported the autofluorescence level to be lower in tumour than in normal brain tissue, but they did not specify the type of brain tissue to be white or grey matter. The fluorescence intensity at 300/395 nm in tumour could correlate with the presence of collagen and was statistically significantly 57% higher from tumour than that from grey matter. The level of collagen in normal adult brain is low, whereas in glioma, non-fibrillar extracellular matrix collagen levels have been demonstrated to increase and might play an important role in driving tumour progression [47]. Some of the fluorescence signal could reflect fibrillar collagen content located around proliferated vasculature.
NADH and FAD are both involved in energy metabolism of cells. Reduced NADH and oxidized FAD are fluorescent, while oxidised NADH and reduced FAD are non-fluorescent [48]. During carcinogenesis, cellular metabolism is often enhanced and tumour tissue would theoretically show increased NADH and decreased FAD [49]. In our study, no significant difference was observed at 340/470 nm (NADH) between grey matter and tumour (Supplement 1, Figure S2). This is possibly because high absorption of blood in the tissue at 470-550 nm hampered the intrinsic fluorescence of this parameters.
There is no significant different of fluorescence intensity at 395/530 nm, which could be ascribed to FAD, between grey matter and tumour ( Fig. 9(b)). However, our study showed the signal on FAD (395/530 nm) from necrotic tissue was statistically significantly 90% lower than that from tumour tissue (Supplement 1, Figure S3). Necrosis typically develops in high-grade gliomas when tumour growth exceeds vascular supply. When tumour cells become ischaemic i.e. are deprived of an oxygen supply and become necrotic, FAD reaches the lowest signal [50]. Though the sample number in this study is limited, the results provide the evidence that necrotic tissue could be differentiated from tumour tissue on the basis of their unique optical properties (Supplement 1, Table S1). The presence of predominantly necrotic tissue for brain needle biopsy is not desirable as it may limit diagnosis and provide insufficient material for further tissue analysis to identify prognostic and predictive biomarkers that may guide further therapy. In some cases this may result in the need for a second biopsy procedure. It is possible that in the future a biopsy needle with an inbuilt optical probe could be used to ensure that biopsy samples are not collected in a necrotic area.
It is important to note that tissue fluorescence correlates to not only the concentration of endogenous fluorophores but is also modulated by the absorption and scattering properties of the tissue. In particular, NADPH and NADH are highly sensitive to changes in their microenvironment such as nutrient, oxygen availability and acidosis [49]. Therefore, in ex vivo tissue samples, NAD(P)H signal may be mixed and artificial. Interpreting changes of tissue fluorescence in brain tissue is complex and a large sample number is needed to establish a reference dataset. In addition, aging might be an important factor as increased autofluorescence pigments, for example, lipofuscin accumulates in neurons of older people [51,52]. Moreover, the infiltrative nature of diffuse glioma further complicates this, given that the cell density of some tumours may be very low. However, where localized tumour growth is present on neuroimaging, the currently investigated techniques may be of benefit. This pilot study was undertaken as proof of concept. Future work would desirably involve analysis of the surgical margin intraoperatively.
In the present study we aimed to make next crucial step towards miniaturizing hand-held probe integrated system for tissue recognition. We identified several source/detector wavelengths where the fluorescence intensities differ from tissue types. VIP scores gave a useful measure to select which predicator variables contribute the most to response variables [53]. Figure 10 illustrates a matrix mapping of source/detector wavelengths used in this study. Fluorophores contributing to tissue autofluorescence and classification are labelled in corresponding source/detector wavelength cells.
Some limitations of our study are as follows. Firstly, there was cross talk between a number or source/detector pairs. For example, due to the broad band of LEDs at wavelength 470 nm, 505 nm and 530 nm, the adjacent detection wavelength channel could detect low levels of DRS, rather than autofluorescence. Therefore, several data points of each optical signal profile have been excluded from the statistical analysis. A laser source is currently in development to be integrated into the system. Secondly, as this was a small pilot study, sample numbers were limited. The algorithms developed were based on the current dataset and should be considered as biased. Based on information we collected from current pilot data, a power calculation was performed to get an idea about how many samples would be needed for a statistically significant sample set; to obtain 80% power (two-tailed, 5% significance) approximately 37 samples for each group would Fig. 10. Matrix mapping of light source/detector wavelengths used in this study. Boxes with yellow colour represent reflectance at eight source wavelengths, boxes with peach colour represent fluorescence at different excitation-emission wavelengths, and main endogenous fluorophores in the brain contributing to tissue autofluorescence tissue classification were labelled as black font in corresponding excitation-emission wavelength boxes. NADH shows no significant difference in our study was marked as grey colour.
be needed, to obtain 90% power (two-tailed, 5% significance) approximately 49 samples for each group would be needed. Further work is underway in order to validate these observations and to improve discriminant algorithm in a larger patient population.
In conclusion, the overall results of this ex vivo study are promising and show that multimodality spectroscopy in a single probe approach is effective for discriminating brain tissues on the basis of their unique optical properties. Certain wavelength pairs (corresponding to relevant biochemical markers) were seen to contain the most discriminative information for tissue classification. This means that less optical components are needed for on-going system development. This is an essential step towards development of a cost-effective, miniaturized hand-held probe integrated tool for real time and label-free surgical guidance.