The study was carried out with approval of the Cedars- Sinai Medical Center Institutional Review Board, with informed consent obtained from each patient. Forty two patients diagnosed with glioma were recruited for the study. The patients underwent craniotomy for surgical removal of brain tumor based on clinical indications. During the surgery, as the brain surface and tumor were removed, the neurosurgeon placed a TR-LIFS fiber optic probe (see below for details) on areas of interest (Fig. 19.1) and the brain tissue was spectroscopically investigated. Areas with distinct pathologic features were selected based on the gross visual evaluation by the neurosurgeon. The goal of the measurements was to record from multiple sites in both the tumor and surrounding normal tissue in each patient. To establish fluorescent signals from “normal”
brain as controls, regions of the exposed brain fur- thest from the visible tumor and apparently tumor-free based on the MRI utilized for intra-operative neuro- navigation were recorded as well. No biopsy was obtained from the normal brain tissues due to ethical limitations. Areas identified as tumors based on con- ventional diagnostic methods (e.g. preoperative MRI) and surgeon experience were TR-LIFS interrogated
during the surgical resection. A small biopsy was per- formed at each spectroscopically investigated spot, except at the areas not considered suitable for physical biopsy due to risks posed to patient. Tumor samples with relatively homogenous morphology were cate- gorized as solid tumors, whereas infiltrated normal tissues at the boundary between the tumor and normal brain tissue was categorized as margins.
Instrumentation
Experiments were conducted with an instrumental setup, which allowed for spectrally-resolved fluores- cence lifetime measurements (Fig. 19.1). A detailed account of this apparatus and its performance has been previously reported (Fang et al., 2004) and will be briefly reviewed here.
Delivery Catheter
Light delivery and collection were implemented with a custom made bifurcated sterilizable probe. It had a central excitation fiber of 600 μm core diameter, surrounded by a collection ring of twelve 200 μm core diameter fibers. The probe was flexible through- out its entire length (3 m) except of a 7 cm distal part consisting of a rigid stainless steel tube. After tissue
Fig. 19.1 Schematic of prototype the Time-Resolved Laser Induced Fluorescence Spectroscopy (TR-LIFS) apparatus. The probe is held over normal cortex to record the Time-resolved
fluorescence
excitation, the emitted fluorescence light was collected and directed into the entrance slit of the spectrometer via the collection channel of the probe. The signal was then detected, amplified, and finally digitized at 8 bits resolution by a digital oscilloscope. The overall time resolution of the systems was approximately 300 ps.
Based on the optical properties of human brain and tumor tissue, we have estimated a penetration depth between ∼250–400 μm for astrocytoma and normal cortex respectively at a wavelength of 337 nm.
Collection of Fluorescent Data – Technical Details The fiber optic probe was positioned 3 mm above the exposed brain tissue specimen with the help of a spacer to optimize the collection efficiency of the probe as previously reported (Papaioannou et al.,2004), the spacer also steadied the probe over the tissue, thus avoiding artifacts in the fluorescence emission due to pulsation of brain. Time-resolved emission of each sample was recorded in the 360–550 nm spectral range and scanned at 10 nm intervals. The energy output of the laser (at the tip of the fiber) for sample excitation was adjusted to 3.0μJ/pulse. The area illuminated by the probe was 2.654 mm2, thus the total fluence per pulse received by the tissue is 1.39μJ/mm2well within safety limits.
Histopathological Analysis
Each biopsy sample was fixed with 10% buffered formalin and stained with Hemotoxylin and Eosin (H&E). Each specimen was independently inter- preted by a neuropathologist with no knowledge of the spectroscopic data. Pathology was the corre- lated with fluorescence spectroscopy measurements for subsequent development of the classification algo- rithm. For the purpose of spectroscopic classifica- tion gliomas were grouped as low grade glioma (LGG-WHO Grade I & II) and high grade glioma (HGG-WHO Grade III & IV). LGG were further subclassified into distinct pathological entities includ- ing oligodendroglioma, oligodendroastrocytoma, and diffuse astrocytoma while high grade gliomas were subcalssified into anaplastic astrocytoma, anaplastic oligodendroglioma, anaplastic oligoastrocytoma and glioblastoma multiforme (Louis et al.,2007).
TR-LIFS Data Analysis
In the context of TR-LIFS, the intrinsic fluorescence impulse response functions (IRF), h(n), describes the real dynamics of the fluorescence decay. The IRF were recovered by numerical deconvolution of the measured input laser pulse from the measured fluo- rescence response transients. The Laguerre expansion technique (Jo et al., 2004a,b) was used for deconvo- lution due to several technical considerations (Butte et al.,2010). This method allows a direct recovery of the intrinsic properties of a dynamic system from the experimental input-output data. The technique uses the orthonormal Laguerre functions to expand the IRF and to estimate the Laguerre expansion coefficients (LEC).
The fluorescence decay profile describes the biochem- ical and morphological characteristics of the tissue.
Each Laguerre coefficient describes various dynam- ics of a complex fluorescence decay curve, i.e. lower order functions describe slower decay characteristics and higher order Laguerre functions describe the faster decay characteristics. The coefficients of the Laguerre function describe the relative contribution of each Laguerre function to the observed fluorescence decay.
In order to characterize and model complete temporal dynamics, all the Laguerre coefficients were computed and used as parameters along with the average life- time which indicates a single value on a complex fluorescence decay curve.
Parameters Selection
Once the fluorescence IRF’s were estimated for each emission wavelength, the steady-state spectrum (Iλ), was computed by integrating each intensity decay curve as a function of time. In order to derive parame- ters from the intensity values, normalized fluorescence spectra was obtained by dividing the discrete intensity values with the intensity value at the peak emission.
It has been shown that blood absorption affects all the wavelengths equally and using intensity ratios negated the requirement for spectral correction (Andersson- Engels et al.,1990). Thus, instead of using normalized intensity values, we used the intensity ratios, which only describe the relative intensity between various spectral bands. Further, to characterize the temporal dynamics of the fluorescence decay, two sets of param- eters were used: (1) the average lifetime (τλ) computed
as the interpolated time at which the IRF decays to 1/eth of its maximum value; and (2) the normalized value of the corresponding LECs. Thus, a complete description of fluorescence from each sample as a function of emission wavelength, Eλ, was given by the variation of a set of spectroscopic parameters at distinct wavelengths (emission intensity – Iλ, average lifetime of fluorescence emission –τλ, and Laguerre coefficients LECλ).
Data Reduction and Statistical Analysis
To identify a set of spectroscopic parameters that best discriminate between various tissue types, a univari- ate statistical analysis (one-way ANOVA) was used to compare the spectroscopic parameters (Iλ,τλ, and LECs) at every Eλ for each type of tissue defined by histology. A p-value of <0.05 was assumed to indi- cate statistical significance. The variables were tested for statistical significance amongst various combina- tions of two tissue types (Fig. 19.2). The significant variables are grouped in six separate sets based on their ability to discriminate between these two types.
Second, as these parameters were used in linear dis- criminant analysis, Lilliefor’s test was performed to ensure the parameters used in the classification were normally distributed. Although, discarding the param- eters based on their non-normality reduces the number of parameters and adds a confounding factor, using nonnormal parameters with linear discriminant func- tions analysis can lead to misclassification (Baron, 1991; Bello,1992).
Classification
In order to classify and predict the fluorescence sig- nal acquired we adopted a classification algorithm based on the binary separation of various tissue types using linear discriminant function analysis (DFA). This approach is a novel application designed to identify a single sample intra-operatively in near real-time (Fig.19.2).
Training phase: Fluorescence emission data was divided in four groups: normal cortex (NC), normal white matter (NWM), low grade glioma (LGG) and high grade gioma (HGG). Since no specific parame- ter was found useful in discriminating all tissues types,
the discriminant model was divided in six models of binary sets, each parsing and classifying all the data set in two tissue types either (normal NC vs. LGG or, NC vs. HGG or, LGG vs. HGG, etc.). For every binary set, the parameters were selected independently. ANOVA was performed to choose the parameters, which were statistically significant between the two types of tis- sues. In order to test for normal distribution, Lilliefor’s test was performed and the parameters with nonnor- mal distribution were discarded. Once the parameters for each set of classifying models were determined, DFA coefficients were calculated and the value of the centroid noted.
Test Phase: A step-wise linear discriminant analy- sis was used to classify tumor tissue (NC, NW, LGG, and HGG) into one of two-group models (e.g. NC ver- sus NWM) based on a set of features (parameters) that describe the tissue and that were derived in the training phase. In general, this approach determined the combination of predictor variables that account for most of the differences observed between the groups.
For instance, the larger the difference between the means of two tissue types relative to the variability within each tissue group, the better the discrimina- tion between the two groups. Assuming that the groups are linearly separable, that is, the groups can be sepa- rated by a linear (or higher dimensional) combination of features (parameters) that describe the tissues being compared. The classification criterion is to assign a tis- sue sample to the group with the highest conditional probability (e.g., Bayes’ Rule); note that this rule also minimizes total error of classification. In general, we are interested in the probability P(Gi/x) that a tissue type belongs to group i, given a set of measurements x.
By applying Bayes’ Theorem we can describe the pos- terior distribution as follows: P(Gi/x)=P(x/Gi)P(Gi), where P(x/Gi) describes the probability of getting a particular set of measurements x given that the tis- sue comes from group i. Prior probability P(Gi) is the probability about the group i known without mak- ing any measurement. Here, we assumed that the prior probability is equal for all. The posterior distribution was used to determine group membership for all six two-group tissues being compared. The total number of posterior probabilities obtained from these models for each tissue type is n–1, where n is the number of tissue types being evaluated. The posterior probabil- ities were then averaged and compared to determine the group to which the sample belongs. In order to
TRAINING PHASE
TEST PHASE
Test Data
Training Set Newly Acquired data
(x)
Cross Validation
Assign tissue to the group with the maximum Discriminant Function (in each of the six group comparisons) Normal Cortex
(NC)
Normal Distribution
Normal Distribution
Normal Distribution
Normal Distribution
Normal Distribution
Normal Distribution NC vs NWM
Iλ, τf , LECs
P(GNC⏐x) P(GNWM⏐x)
P(GNC⏐x)
Σ n
P(GNC⏐x) P(GLGG⏐x) P(GNC⏐x) P(GHGG⏐x) P(GNWM⏐x) P(GLGG⏐x) P(GNWM⏐x) P(GHGG⏐x) P(GLGG⏐x) P(GHGG⏐x)
SAVE SIX SETS OF TRANING PARAMETERS
Iλ, τf , LECs Iλ, τf , LECs Iλ, τf , LECs Iλ, τf , LECs Iλ, τf , LECs
ANOVA ANOVA ANOVA ANOVA ANOVA ANOVA
NC vs LGG NC vs HGG NWM vs LGG NWM vs HGG LGG vs HGG
NC vs NWM NC vs LGG NC vs HGG NWM vs LGG NWM vs HGG LGG vs HGG
Linear Discriminant Analysis Posterior Probabilities Estimate Group Membership Tissue Types
Binary Sets
Determine significant parameters within a set Lilliefors test for Normality Extracted Spectroscopic Parameters Normal White
Matter (NWM)
Low Grade Glioma (LGG)
High Grade Glioma (HGG)
P(GNWM⏐x)
Σ n ΣP(GLGGn ⏐x) ΣP(GHGGn ⏐x)
Fig. 19.2 Elements of classification algorithm to identify a single sample time-resolved fluorescence spectroscopy data acquired intra-operatively. The posterior probability of the
sample belonging to the groups being analyzed is calculated using linear discriminant analysis. All the posterior probabilities are averaged for each type of tissue to determine the diagnosis eliminate bias in sensitivity and specificity due to the
data driven predictions, a leave-one- out cross valida- tion approach was used to calculate the classification (discriminant) score and thereby predict group mem- bership. In the leave-one-out cross-validation, a single set of spectroscopic data is removed from the study population and used in the validation process. The cutoff value was evaluated in a data-driven way from the remaining n–1 parameters. Thereafter, the resulting cutoff value was applied to the newly acquired data.
This tissue was then classified as a true positive, a false positive, a false negative, or a true negative, depending on whether this tissue is classified as belonging or not to the group membership being assessed. This process was repeated for all parameters in the dataset, and the results based on all of the parameters then used to eval- uate the sensitivity and specificity corresponding to the cutoff value that was derived in the n–1 parameters.
Results
Fluorescence emission data were collected from a total of 186 samples. Pathological analysis of the tissue corresponding to these measurements was available in 71 samples. These include normal cortex (N = 35);
normal white matter (N = 12); low grade glioma (N=7) including mixed oligoastrocytoma and oligo- dendroglioma (N =4) and diffuse astrocytoma (N = 3); and high grade glioma (N=17).
Time-Resolved Fluorescence Characteristics
All samples (Fig. 19.3a, b) showed relatively broad emission spectrum with a varying degree of reduced emission at 415 nm which corresponds to the band
Fig. 19.3 Comparison of average fluorescence values mean± SE of the spectroscopic parameters, emission spectra intensity, average lifetime, average Laguerre coefficients (LEC-0, LEC-1)
across the emission wavelengths for distinct brain tissue types (a) normal cortex, normal white matter and low grade glioma, (b) normal cortex, normal white matter and high grade glioma
of hemoglobin absorption (Beychok et al., 1967).
Variability of the signal at 415 nm was due to differ- ent amounts of blood interfering with the optical probe during surgery. Theτλ values follows a similar trend for all tissue types; lifetime at blue-shifted wavelengths (∼390 nm) was typically longer when compared with red-shifted wavelengths (> 440 nm). LECs as a func- tion of wavelengths provided additional information to fluorescence spectra and lifetime for comparing and classifying tissue samples. The zeroth-order coefficient (LEC-0) and the first order (LEC-1) were used to clas- sify the tissues. LEC-0 closely followed the average lifetime in data, whereas LEC-1 provided information regarding faster dynamics in the fluorescence decay curve and thus was used as an additional parameter for tissue characterization.
Normal Cortex (NC): NC fluorescence was char- acterized by a broad fluorescence emission with two distinct peaks centered at 390 and 440 nm, with the emission slightly higher at 440 nm. Theτλ values at 390 nm (τ390=2.12±0.10 ns) were longer than those at 440 nm (τ460=1.16±0.08 ns). The LEC-1 values at 390 nm (LEC-1390=–0.032±0.019) were lower
when compared with LEC-1 values at 440 nm (LEC- 1440 =0.095 ±0.010), a trend that mirrored theτλ values.
Normal White Matter (NWM): The NWM spectra showed a main peak emission shifted to 450 nm when compared to NC. The emission at 390 nm (I390=0.63
± 0.07) was also lower when compared to NC. The average lifetimes at 390 nm (τ390=1.933±0.15 ns) and 440 nm (1.193±0.11 ns) were similar to those of NC. The LEG-0 had similar values at 390 nm (LEC- 0390=0.766±0.02) and 440 nm (LEC-0440=0.7269
±0.018) to those on NC. The LEG-1 at 390 nm (LEC- 1390=–0.00885±0.002) also similar to LGG while at 440 nm (LEC-1440 =0.105± 0.020) was similar to NC.
Low Grade Glioma (LGG): LGG demonstrated a main peak emission centered at 440–450 nm. A sec- ondary peak was present around 390 nm, however, this was lower (∼30%) when compared with the main peak emission (Fig.19.3C). Theτλvalues appeared slighter longer than those observed in NC, both at 390 nm (τ390 = 1.81 ± 0.35 ns) and 440 nm (τ440 = 1.38 ± 0.31) but this difference was not statistically
significant. The LEC-0 at 390 nm (LEC-0390=0.7749
±0.34) was similar to 440 nm (LEC-0440 =0.775± 0.33). The LEG-1 at 390 nm (LEC-1390=–0.0077± 0.0045) was higher than the LEC-1 values observed at 440 were lowest in (LEC-1440 =–0.068 ± 0.0041).
The LEC-1440for LGG was the lowest of all observed for all tissue types. Notably, we determine differences within the subsets of LLG. Diffuse astrocytoma was found to have higher emission intensity and longerτλ at 390 nm than oligodendroglioma. When the fluores- cence from oligoastrocytoma and oligodendroglioma were studied separately, theτλat 390 nm was shorter (τ390=1.57±0.2 ns) compared with diffuse astrocy- toma (τ390=2.58±0.5 ns).
High Grade Glioma (HGG): HGG demonstrated a main peak emission centered at 450 nm. A secondary peak was present around 390 nm. This was lower (∼40%) when compared with the 450 nm emission.
Average lifetime observed at 390 nm (τ390 =1.93± 0.18 ns) was longer than at 440 nm (τ440 =1.138± 0.12 ns). The LEG-0 at 390 nm (LEC-0390 = 0.761
± 0.023) was higher than at 440 nm (LEC-0440 = 0.0394 ±0.016). The values of the LEG-1 observed at 390 nm and 440 nm were (LEC-1390 = –0.0183
±0.34, (LEC-1440 =0.0875±0.0223) respectively.
As observed in LGG tumors, HGG demonstrated variation in the fluorescence emission characteristics based on histopathological sub-classification. It was noted that the fluorescence emission data collected from HGG demonstrated a great degree of variability.
Glioblastoma multiforme (GBM) fluorescence emis- sion characteristics were similar to low-grade oligo- dendroglioma with a shorter τλ at 390 nm (τ390 = 2.145 ± 0.25 ns) compared to anaplastic oligoden- droglioma (τ390 =2.66 ± 0.5 ns) with longerτλ at 390 nm. A single sample of recurrent glioma with necrotic changes was characterized by a single emis- sion peak at 390 nm of wavelength and overall faster average lifetime. (τ390 =1.11 ns) fig. 4b. In addition, we observed difference in the fluorescence parameters
of HGG core compared with margins of the tumor within the same patient (data not shown). The fluo- rescence parameters observed for the tumor margins resembled those of LLG.
Statistical Analysis and Classification
Table19.1depicts the classification results using both spectral intensities ratios and time-resolved parame- ters, computed based on the algorithm that allows for binary separation of distinct tissue groups using LDA.
All LGG were correctly identified, except for one mea- surement in which NC was classified as LGG. The accuracy dropped when classifying HGG , with a sen- sitivity of 47% and specificity of 95%. We attribute this lower sensitivity value for HHG to high variability in the fluorescence signature of HGG.
Discussion
TR-LIFS can accurately differentiate LGG from nor- mal tissues (100% sensitivity and 98% specificity).
These finding are in agreement to those reported previ- ously by our group (Butte et al.,2010). Differentiating between LGG and NWM is of importance when try- ing to achieve a complete excision at the margins, as the tumor will be surrounded by NWM. Detection of LGG represents a major challenge to current tumor resection. LGG are often visually bland. As such they are far more difficult to differentiate from surrounding normal brain than HGG. Further, these tumors typi- cally infiltrate into surrounding tissues that maintain function, often resulting in surgeons performing only biopsies or limited resections. Several recent studies have demonstrated that the extent of resection of LGG correlates with long term survival (Sanai and Berger, 2008; Smith et al., 2008). Consequently, one of the
Table 19.1 Classification results using TR-LIFS TR-LIFS
Histopathology NC NWM LGG HGG Sensitivity (%) Specificity (%)
NC (n=35) 27 3 1 3 79.41 78.26
NWM (n=12) 4 8 0 0 66.67 90.63
LGG (n=7) 0 0 7 0 100.00 98.44
HGG (n=17) 6 3 0 8 47.06 94.64
great potential applications of TR-LIFS is the ability to differentiate LGG from normal brain as an adjunctive tool for increasing the extent of resection.
The HHG were classified with high specificity (95%) but very low sensitivity (47%). Although, high specificity reduces the risk of resecting normal brain, we attribute this to the high variability in the TR-LIFS signals obtained from various subclasses of HGG. We noted these tumors exhibited highly heterogeneous fluorescence spectroscopic features. Such differences may be attributed to the large variance in the pro- tein expression (Umesh et al.,2009) within the same HGG tumor. Generally, HGG cells are more pleomor- phic when compared with LGG. Thus, future studies will need to determine how well heterogeneities in HHG can be distinguished using TR-LIFS derived parameters. We anticipate that a more comprehensive classification based on the biochemical and immuno- histochemical features will improve the classification accuracy.
The small sample size available for this study most likely also contributed to the low sensitivity values.
In order to obtain an unbiased estimation of the clas- sification accuracy, a larger set of data needs to be obtained and the training set and the test set used have to be completely independent. Consequently, a higher number of HHG need to be investigated along with the biochemical heterogeneities within the same tumor for a more comprehensive assessment of the ability of TR-LIFS to distinguish HGG tissue from normal brain tissue.
The NC measurements were obtained from areas distal to the tumor site and, where the arachnoid and pia covering the cortex were generally preserved.
Arachnoid and pia mater consists predominately of collagen fibers. Collagen has peak fluorescence at 390 nm of wavelength with an average lifetime of
∼3 ns (Marcu et al.,2001). The change in collagen amount between optical probe and the actual brain tissue may have contributed to the variation in the aver- aged lifetime values. Due to our inability to perform a biopsy on the normal healthy brain, we were unable to confirm this hypothesis.
We note that the average lifetime values at 390 nm (1.45±0.4 ns) of LGG in our previous study (Butte et al.,2010) were found shorter than those observed in the current study (1.58 ± 0.2 ns). To understand this difference in average lifetime value at 390 nm from LGG samples, we subdivided the LGG tumors
into mixed astrocytoma, oligodendroglioma and dif- fuse astrocytomas. It was observed that the oligo- dendral tumors had faster lifetime at 390 nm com- pared to tumors of astrocytic origin such as diffuse astrocytomas. This is a significant finding as diffuse astrocytomas have a tendency to convert to GBM We hypothesize that this difference may be attributed to the IDH-1 mutation in oligodendral tumors (Hartmann et al., 2009; Yan et al., 2009), which affects the NADP+ dependant isocitrate dehydrogenase leading to an up-regulation of glutamate decarboxylase (GAD) which has peak fluorescence emission at 390 nm of wavelength and average lifetime of 1.8–2.1 ns (Rosato et al., 1989). We are in the process of designing new experiments to confirm presence of both colla- gen and Glutamate decarboxylase as the fluorophores responsible for this variation.
The classification using only the spectral intensity ratios demonstrated a significant drop in the sen- sitivity and specificity of normal white matter and high grade glioma, whereas, time-resolved fluores- cence data without any correction for spectral response yielded superior accuracy. This emphasizes the poten- tial of time resolved fluorescence in the intra-operative setting in which factors which affect the spectral intensity values such as hydration, temperature, and absorption by blood cannot be controlled. The flu- orescence decay characteristics are immune to these factors.
Limitations of classification method: While we were able to acquire large fluorescence data that allows for the formation of two data sets training and test, we were not able to include all data in the classification stage. We were limited by constrains on the number of biopsies performed on the brain tissue. Only a limited number of spectroscopic data sets can be correlated with the histopathological diagnosis. Consequently, we used a cross-validation method (leave-one-out) to test our algorithm. Although cross-validation method can remove the bias in such analyses, there is a danger in overestimation of the data by selecting the best possi- ble outcome. In addition, we used known pathology to determine the best set of parameters to be used in the classification. Such assumption in the analysis can lead to distorted and predetermined results (Kriegeskorte et al.,2009). In order to avoid this issue we anticipate acquiring more data with histopathological analysis in the future to create a separate training set and test set.