1.

Introduction

According to the Epilepsy Foundation, of the 200,000 new cases of epilepsy diagnosed each year in the United States, 45,000 are in children under the age of 15.¹ Epilepsy is considered a severe neurological disorder in children, because its primary symptom, seizure, can severely damage normal brain development. When pharmaceutical and/or dietary therapies fail to achieve seizure freedom, epilepsy is defined as refractory.^2,3 For pediatric patients with refractive epilepsy, surgical intervention, when applicable, becomes necessary. Since the success of epilepsy surgery is determined by the amount of seizure-inducing cortex (i.e., epileptic cortex) removed, there arises a need to develop some systematic methodology that accurately detects and demarcates the epileptic cortex pre- and intra-operatively.

Noninvasive diagnostic technologies commonly used in the presurgical planning of epilepsy surgery include electroencephalography (EEG), video EEG, and several noninvasive imaging modalities like computed tomography (CT), single-photon emission computed tomography (SPECT), magnetic resonance imaging, positron emission tomography, and diffusion tensor imaging.^4–8 Advancements in noninvasive imaging modalities over the past two decades have greatly improved their ability to identify epileptic cortex or lesions and hence improve surgical outcomes.^4,7,8 However, the above mentioned diagnostic technologies, when used individually, cannot provide definitive information on which surgery can be based; at best, they provide approximate locations of seizure origins, due to their limited sensitivity and spatial resolution.^4,7,9 In addition, these technologies may be difficult to use on pediatric patients, who may not be able to remain motionless for extended periods.¹⁰ Finally, spatial coordinates of the epileptic cortex defined in these studies may lose their value because of brain shift/deformation during the craniotomy procedure.^11–14

When EEG/imaging studies are inconclusive, electrocorticography (ECoG) may be used to perform invasive mapping of the ictal onset zone.^9,15 In addition, ECoG is used in the operating room to detect interictal epileptiform discharges directly from the abnormal brain region; these discharges are used to define the resection margins for the surgeon. However, ECoG may also detect interictal discharges in a normal brain, as a result of electrical spreading, and epileptic focus localization using interictal discharges is subject to a certain amount of uncertainty in an anesthetized brain.¹⁵ Histological evaluation of the epileptic cortex has uncovered many unique pathological features, such as balloon cells and cortical dislamination.¹⁶ This information usually is acquired through postoperational histopathological analysis of the resected brain specimens and is not immediately available in the operating room. While histopathological features of the lesion may not correlate with surgical outcomes,^7,17 it is not yet clear if their presence can be used as an objective indicator with which to demarcate the area of an epileptic brain. The technological shortcomings and limitations described above frequently lead to incomplete removal of epileptic lesions, which adversely affects surgical outcomes. This is especially true for cases of extra-temporal epilepsy, which is common in children.^4,8,10,17 For these reasons, new technologies for intraoperative identification and demarcation of the epileptic brain area are needed to enhance the prognosis of epilepsy surgery for patients with intractable nonlesional extra-temporal epilepsy. These new technologies, ideally, should be developed in accordance with the intrinsic characteristics of an epileptic brain during the interictal period.

Because the intrinsic compositional and structural characteristics of tissue strongly affect its absorption and scattering properties, it is feasible to use the interaction between light and biological tissue to estimate these qualities and hence detect whatever alterations that result from disease development.^18–28 Such an optic-based diagnostic technology often is termed optical diagnosis; its advantages include its low cost, portability, and real-time application. Many of the existing optical diagnostic methodologies are based on fluorescence and diffuse reflectance spectroscopy. This is because several intrinsic biological chromophores (e.g., hemoglobin) and fluorophores (e.g., NADH) can be used to assess critical intrinsic physiological characteristics of biological tissue, including its hemodynamics and metabolism.^{26,27,29–31} The feasibility of using optical spectroscopy to guide cancer surgery intra-operatively has been evaluated for various cancer types. Several studies have shown that combined fluorescence and/or diffuse reflectance spectroscopy can accurately detect brain tumors in vivo and hence may be used to guide brain tumor resection.^21,32 However, applying optical spectroscopy to the detection of the epileptic cortex has been explored in only a limited fashion to date.^33–37

In this paper, combined fluorescence and diffuse reflectance spectroscopy is used to detect unique pathophysiological and morphological characteristics of the epileptic cortex in an in vivo human study. Static diffuse reflectance and fluorescence spectra are acquired from in vivo epileptic and normal cortex of pediatric patients undergoing epilepsy surgery. Spectral data are analyzed using various statistical tools, and unique spectral features associated with epileptic brain tissue defined by histology and ECoG are identified. The physiological underlying mechanisms responsible for these spectral alterations are extrapolated.

2.

Materials and Methods

2.4.

Data Processing

All spectral data acquired from the in vivo study were preprocessed to remove any artifacts induced by instrumentation. Mainly, baseline subtraction and calibration factor applications were performed to all acquired fluorescence spectra $F(\lambda )$ and diffuse reflectance spectra $Rd(\lambda )$. The calibration factors used here were derived from the ratio of true and measured emission spectra from a calibrated light source.³⁸ Furthermore, to speed up the data analysis procedure, spectral range and resolution were further reduced. Specifically, diffuse reflectance data between 400 and 850 nm and fluorescence data between 380 and 750 nm were preserved; spectral resolution was reduced to 5 nm. In addition to the individual spectral types, a combined spectrum was created for the data analysis process, specifically the ratio of the fluorescence and diffuse reflectance spectra [i.e., $F(\lambda )/Rd(\lambda )$]. The wavelength range of the combined spectra was 400 to 750 nm, and the resolution was 5 nm. Several normalization techniques were applied to the preprocessed spectra, including local maximum normalization, area-under-the-curve normalization, global mean normalization, and arbitrary reference normalization (Fig. 1). For local maximum normalization, each spectrum is normalized to its maximum value. That is, $S_{\max }(\lambda )=S(\lambda )/\mathrm{MAX}[S(\lambda )]$, where $S(\lambda )$ represents either $F(\lambda )$, $Rd(\lambda )$, or $F(\lambda )/Rd(\lambda )$, and MAX is the maximum operation. For area-under-the-curve normalization, the new spectrum $S_{\text{AREA}}(\lambda )$ is calculated by $S_{\text{AREA}}(\lambda )=S(\lambda )/\int S(\lambda )d\lambda$. For global mean normalization, each spectrum is normalized to the mean of the total spectral data set. That is, $S_{\text{MEAN}}(\lambda )=S(\lambda )/[\sum _n{S}_i(\lambda )/n]$, where $n$ represents the total number of spectra in the dataset. Finally, arbitrary reference normalization produces a new spectrum using $S_{\text{POINT}}(\lambda )=S(\lambda )/S({\lambda }_{\mathrm{ref}})$, where ${\lambda }_{\mathrm{ref}}$ is an arbitrary reference wavelength.

Fig. 1

The flow chart of the data processing procedure employed in the study.

3.

Results

A total of 16 patients were enrolled in this study; their demographic information is summarized in Table 1. A total of 84 useful investigated sites were obtained from these patients: 28 located outside the resection zone defined by ECoG (i.e., normal brain) and 56 within the zone of resection. The 56 sites within the zone of resection were divided into two subcategories in accordance with the results of our histological evaluation; 30 sites exhibited histological features associated with epilepsy (epileptic brain), while the other 26 sites did not. The distribution of sites is shown in Table 2.

Table 1

Demographic information of the patients studied.

Table 2

Distribution of the investigated sites.

Spectral data analysis revealed that only those spectral differences identified within the original $Rd(\lambda )$, $R{d}_{\mathrm{MEAN}}(\lambda )$, $R{d}_{\text{POINT}}(\lambda )$, and $F/R{d}_{\text{POINT}}(\lambda )$ could produce discrimination algorithms with a sensitivity and specificity greater than 75%. Analytical results for each spectral type follow.

3.1.

Original $Rd(\lambda )$

Comparing the original $Rd(\lambda )$ from normal and epileptic brains revealed significantly different mean $Rd(\lambda )$ intensities between 400 and 600 nm ($p<0.05$, see Fig. 2). Inputting the $Rd$ signals from this spectral band, the best classification result based on LDA was 82.1% sensitivity, 76.7% specificity, and 79.3% total accuracy.

Fig. 2

(a) The solid black line represents the mean original $Rd(\lambda )$ from the normal brain group, and the solid pink line represents the mean original $Rd(\lambda )$ from the epileptic brain group. The error bars represent one standard deviation of the mean. (b) Statistical mean comparison of original $Rd(\lambda )$ between normal and epileptic brains; $p$ value is below 0.05 between 400 and 600 nm. The optimal performance of the classification based on LDA, using original $Rd(\lambda )$ as the inputs, was $\text{sensitivity}=82.1\%$, $\text{specificity}=76.7\%$, and total correct $\text{classification}=79.3\%$.

3.2.

$R{d}_{MEAN}(\lambda )$

Similar to original $Rd(\lambda )$, $R{d}_{MEAN}(\lambda )$ between 400 and 600 nm also was statistically different between normal and epileptic brains ($p<0.05$, see Fig. 3). Entering the signals from this spectral band, the best classification result based on LDA was 75.0% sensitivity, 83.3% specificity, and 79.3% overall accuracy, which is comparable to using the original $Rd(\lambda )$.

Fig. 3

(a) The solid black line represents the mean ${Rd}_{\mathrm{MEAN}}(\lambda )$ from the normal brain group, and the solid pink line represents the mean ${Rd}_{\mathrm{MEAN}}(\lambda )$ from the epileptic brain group. The error bars represent one standard deviation of the mean. (b) Statistical mean comparison of ${Rd}_{\mathrm{MEAN}}(\lambda )$ between normal and epileptic brains; $p$ value is below 0.05 between 400 and 600 nm. The optimal performance of the classification based on LDA, using ${Rd}_{\mathrm{MEAN}}(\lambda )$ as the inputs, was $\text{sensitivity}=75\%$, $\text{specificity}=83.3\%$, and total correct $\text{classification}=79.3\%$.

3.3.

$Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$

Analyzing $Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$, as shown in Fig. 4, revealed several spectral features that could effectively differentiate a normal brain from an epileptic brain. Most of these spectral features were evident between ${\lambda }_{\mathrm{ref}}=400$ and ${\lambda }_{\mathrm{ref}}=600\mathrm{nm}$. Entering these identified spectral features, classification results based on LDA were $\sim 75\%$ sensitivity and $\sim 80\%$ specificity (Table 3). Discrimination was optimized when ${\lambda }_{\mathrm{ref}}$ was 505 nm, where the overall accuracy of the discrimination algorithm was 86.2%.

Fig. 4

(a) Surface plot of the $p$ values from the mean comparison of $Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$ between normal and epileptic brains. (b) Representative test results with reference wavelength ${\lambda }_{\mathrm{ref}}=505\mathrm{nm}$. The solid green line represents the mean $Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$ of the normal brain group, and the solid blue line represents the mean $Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$ of the epileptic brain group. Statistically significant differences (i.e., $p<0.05$) are found between 425 and 460 nm and between 535 and 575 nm in this case.

Table 3

Classification results of the discriminations based on LDA, using Rd(λ)/Rd(λref) as the inputs.

${\lambda }_{\mathrm{ref}}$ (nm)	Sensitivity (%)	Specificity (%)	Total correct classification (%)
445	75	86.7	81
455	75	80	77.6
460	75	83.3	79.3
500	78.6	90	84.5
$\mathbf{505}$	82.2	90	86.2
535	78.6	86.7	82.8
550	78.6	86.7	82.8
605	78.6	86.7	82.8

The bold values represent the best results.

3.4.

$[F(\lambda )/Rd(\lambda )]/[F({\lambda }_{\mathrm{ref}})/Rd({\lambda }_{\mathrm{ref}})]$

Analysis of $[F(\lambda )/Rd(\lambda )]/[F({\lambda }_{\mathrm{ref}})/Rd({\lambda }_{\mathrm{ref}})]$, as shown in Fig. 5, also revealed several spectral features from the combined spectrum that could effectively distinguish a normal brain from an epileptic brain. Again, the majority of these spectral features were located between ${\lambda }_{\mathrm{ref}}=400$ and ${\lambda }_{\mathrm{ref}}=600\mathrm{nm}$. Entering the identified spectral features, the classification results based on LDA yielded a sensitivity above 75% and a specificity above 80% (Table 4). Discrimination was optimized when ${\lambda }_{\mathrm{ref}}$ was 485 nm, where the overall accuracy of the discrimination algorithm was 96.6%.

Fig. 5

(a) Surface plot of $p$ values from the mean comparison of $[F(\lambda )/Rd(\lambda )]/[F({\lambda }_{\mathrm{ref}})/Rd({\lambda }_{\mathrm{ref}})]$ between normal and epileptic brains. (b) Representative test results with reference wavelength ${\lambda }_{\mathrm{ref}}=485\mathrm{nm}$. The solid green line represents the mean $[F(\lambda )/Rd(\lambda )]/[F({\lambda }_{\mathrm{ref}})/Rd({\lambda }_{\mathrm{ref}})]$ of the normal brain group, and the solid blue line represents the mean $[F(\lambda )/Rd(\lambda )]/[F({\lambda }_{\mathrm{ref}})/Rd({\lambda }_{\mathrm{ref}})]$ of the epileptic brain group. The statistically significant differences (i.e., $p<0.05$) are found between 430 and 450 nm and between 625 and 660 nm in this case.

Table 4

Classification results of the discriminations based on LDA, using [F(λ)/Rd(λ)]/[F(λref)/Rd(λref)] as the inputs.

${\lambda }_{\mathrm{ref}}$ (nm)	Sensitivity (%)	Specificity (%)	Total correct classification (%)
400	75.0	83.3	79.3
420	75.0	86.7	81.0
455	89.3	86.7	87.9
$\mathbf{485}$	92.9	100.0	96.6
500	85.7	90.0	87.9
505	89.3	90.0	89.7
510	89.3	90.0	89.7
520	92.9	90.0	91.4

The bold values represent the best results.

4.

Discussion

The in vivo study results demonstrate that in vivo static diffuse reflectance spectroscopy, as well as its combination with static fluorescence spectroscopy, can be used effectively to differentiate an epileptic brain (as defined by ECoG and histology) from a normal brain with a high degree of accuracy. The most effective discrimination algorithm is derived using the spectral features from $[F(\lambda )/Rd(\lambda )]/[F(485\mathrm{nm}/Rd(485\mathrm{nm})]$, where discrimination accuracy reaches 96.6%. This supports the concept of developing a new intraoperative guidance system, based on optical spectroscopy, to detect histopathological abnormalities associated with an epileptic brain in the operating room and thereby provide additional information to aid in the demarcation of epileptic foci.

Using parametric and nonparametric statistical comparison methods, original $Rd(\lambda )$, $R{d}_{\mathrm{MEAN}}(\lambda )$, and $Rd(\lambda )/Rd({\lambda }_{\mathrm{ref}})$ between 400 and 600 nm were found to be significantly different in epileptic brains versus normal brains. Since diffuse reflectance spectra in this region are heavily influenced by hemoglobin absorption, this finding implies that certain static hemodynamic characteristics of the epileptic cortex, in particular hemoglobin concentration and hemoglobin oxygenation, are very different from that of a normal cortex. Previous studies have shown that hypometabolism and hypoperfusion are frequently observed in the area of an epileptic focus and its immediate surroundings in the brain of patients with partial seizures.^39–44 In addition, Stefanovic et al. pointed out that normal interictal neurovascular coupling is progressively compromised by sustained seizure attacks.⁴⁴ Schwartz suggested that the epileptic cortex generally receives inadequate cerebral blood flow.⁴⁵ However, alterations in resting state hemodynamics may not exist in all types of refractive epilepsy. It has been shown that the average resting metabolism and blood flow of the epileptic cortex are very similar to those of the normal cortex in patients with idiopathic generalized epilepsy.^46–48 Because of significant differences in the discharge characteristics of ictal versus interictal states, it has been suggested that the cerebral blood flow and cerebral metabolic rate of oxygen changes induced by interictal discharges may be far less noticeable than their ictal counterparts.^39,47,49 This implies that the changes in diffuse reflectance signals reported here are not necessarily associated with the presence of interictal discharges. The results of this study further suggest that histological abnormalities may be linked to changes in regional resting state hemodynamics. It is possible that histological abnormalities like cortical dysplasia modify the characteristics of astrocytic ${\mathrm{Ca}}^{2+}$ oscillations, which in turn alter astrocytic function regulating cerebral blood vessel diameters and hence regional hemodynamics.^50–54 More studies must be conducted to verify such conjectures.

Several studies have shown that the metabolic activities of any in vivo biological tissue, including the brain, may be monitored through the redox state of NAD(P)H and hence its fluorescence.^27,28 Since the epileptic cortex is hypometabolic,^39–44 we hypothesized that fluorescence spectroscopy alone may be used to detect this characteristic of an epileptic brain. However, the results reported here do not support this hypothesis; no statistically significant differences were found between the fluorescence spectra from an epileptic brain and those from a normal brain. One possible explanation for this inefficiency is that NAD(P)H fluorescence signals measured from an in vivo brain are heavily modulated by brain absorption (i.e., hemoglobin concentration and hemoglobin oxygenation) and scattering properties. Under this condition, complex spectral processing techniques⁵⁵ would have to be used in conjunction with fluorescence spectroscopy to recover intrinsic NAD(P)H fluorescence quantity and hence detect the hypometabolic characteristic of the epileptic cortex.

The results of this study are promising but preliminary in nature. Several obstacles must be overcome before the clinical value of optical spectroscopy, namely fluorescence and diffuse reflectance spectroscopy, is clarified for demarcating epileptic foci and hence aiding epilepsy surgery. The most important challenge is the lack of a gold standard for defining an epileptic brain. While optical spectroscopy can indirectly detect histological abnormalities of an epileptic brain, it is not yet proven that these histological features provide better demarcation than ECoG. Furthermore, this study also revealed considerable discrepancies between the results of ECoG and those of histological study. On many occasions, the brain area classified as abnormal by ECoG was interpreted as histologically normal by our neuropathologist. This may be attributed to the fact that only small biopsy samples (i.e., less than $10{\mathrm{mm}}^3$) were taken from the investigated sites within the resection zones for the purpose of accomondating the limited depth of investigation of the optical probe. Therefore, histopathological abnormalities present in the deeper layers of the cortex could be missed. Another possible explanation is that conventional histological methods cannot detect molecular alterations underlying the abnormal electrical activities of the epileptic cortex. Also, it is possible that the origin of interictal spikes may not be the same as that defined by ECoG due to electrical spreading. Finally, it will be challenging to establish a universal normal baseline for the diffuse reflectance and fluorescence spectra of a normal pediatric cortex, because the brain undergoes development and increases in size and weight significantly throughout childhood. These issues will be addressed in follow-up studies to be conducted by the current investigators, and their results will be presented in future publications.

5.

Conclusions

The feasibility of using diffuse reflectance and fluorescence spectroscopy to aid epilepsy surgery in children was investigated. Static diffuse reflectance and fluorescence data were collected from the abnormal cortex, as defined by pathology and ECoG. Tissue discrimination algorithms with high sensitivity and specificity were produced using various spectral features. The success of this study warrants a future large-scale study to verify the potential role of optical spectroscopy in pediatric epilepsy surgery.