Human specimens
Human tissue measurements were performed at the Anatomy department of the Radboud University Medical Center Nijmegen. In this study STS measurements were performed in 19 human cadavers, 11 male and 8 female with an age range of 65–92 years (mean age 80.4 years). A handwritten and signed codicil from the donor, as required by the Dutch law for scientific research and education is kept at the Department of Anatomy of the Radboud University Nijmegen, Nijmegen, The Netherlands.
The human cadavers were one-time frozen and selected to have no neurological abnormalities or diseased tissue near the intended measurement sites. In each cadaver the cervical area (from interscalene region to proximal until the neuroforamen) and forearm area (medial side of the mid forearm) was dissected to expose the nerves and surrounding tissues. In all human cadavers cervical nerves and surrounding tissues were measured and in ten of them the median nerves with surrounding tissues were also measured.
Cervical nerve, median nerve and surrounding tissue anatomy
In general after skin perforation, when advancing the needle towards the nerve a layer of subcutaneous fat, muscle tissue and the fat surrounding the nerve (sliding fat) will be traversed (see Fig. 1). Sliding fat in our case is defined as fat surrounding the entire nerve bundle; fat outside the outer epineurium. In Fig. 2 the histology of subcutaneous fat, muscle, the cervical nerve and the median nerve is shown. The cervical nerve at the proximal side is mostly mono-fascicular branching into multiple fascicular bundles at the distal side. The fascicular bundles are surrounded by loose collagenous tissue and adipose tissue of the epineurium [13]. The median nerve is already multi fascicular at the proximal side, with the number of fascicles increasing towards the distal side. Subcutaneous fat shows a high amount of adipose cells with some connective tissue while muscle tissue is mostly composed of muscle fibers.
Instrumentation
Ex vivo diffuse reflectance spectra were acquired using a portable spectroscopic system as illustrated in Fig. 3. A tungsten halogen broadband light source with an integrated shutter (Avantes HAL-S) was used as an emitting source. Delivery of light to the tissue and its collection were achieved with a 1.0 mm diameter fiber-optic probe with the distal end polished at an angle of 30 degrees. The probe consisted of two 200-μm core diameter optical fibers with a tip to tip distance of 0.85 mm. The illumination fiber was connected to the light source and the collection fiber was connected to a spectrometer with a silicon detector (Horiba, S318-2 VIS) and a spectrometer with an InGaAs detector (Horiba-S330-2 NIR), via a 50–50 % fiber splitter to divide the collected light over the two detectors.
The diffuse optical spectra were calibrated by subtracting a background measurement acquired after each tissue measurement and divided by the response of a white reflectance standard (Spectralon) with known reflectivity.
Acquisition time per spectrum was 1 s including background measurement. At each measurement location typically 10 spectra were acquired that were averaged to form one data point in the database.
Data acquisition
In this study STS measurements were performed on 19 human cadavers. A database was compiled of spectra measured at 1274 different locations on and near nerve roots as well as at 289 different locations on and near the median nerve. Spectra (400–1710 nm) were acquired from subcutaneous fat, muscle, sliding fat, and fascicular tissue of the nerve. Two experienced anesthesiologists, with the assistance of an anatomist, positioned the optical needle on the exposed tissues. In all cases we ensured that the optical fibers of the needle made good contact with the tissue to avoid artifacts due to air voids. Because the needle was held by hand some slight movements could have occurred but analysis of several measurements taken at one location did not show significant variations in the spectra. To avoid that a mixture of the tissue types was measured we made sure that the tissue layer was thicker than 2 mm. Especially for sliding fat, the optical needle was placed such that no nerve structure was in the direction of the light beam.
Feature extraction methods
All data processing was performed using software developed in-house based on MATLAB (Mathworks Inc.) and its Statistics toolbox. Where third party toolboxes were used, they are mentioned below.
With a resolution of 1 nm each measured spectrum consists of 1311 data points. To avoid over-fitting, the number of data points must be translated into features, with the number of features considerably lower than the number of spectra measured. Several methods to reduce the number of data points into relevant features were considered. In particular we considered methods with limited dependence on absolute intensities since the absolute intensity of the spectra may fluctuate significantly due to small variations in calibration conditions, the state of the probe, or even the bending radius of the fibers.
The first feature extraction method was a fitting method which decomposes a spectrum into 12 fit parameters i.e. features. The second method was to perform a standard principal component analysis (PCA) and use the most important principal components as features [14].
Physiological fit parameters
Spectral data modelling can relate the measured spectra to underlying physiological parameters like blood and fat concentrations. The spectra were fit with a modified version of the model developed by Farrell et al. [15]. This model estimates the absorption coefficient \(\upmu_{\text{a}} \left(\uplambda \right)\) and the reduced scattering coefficient μs′(λ) expressed in cm−1. To determine the chromophore concentrations present in the tissue, the fiber distance between the emitting and collecting fibers as well as the wavelength-dependent absorption coefficients of these chromophores need to be known [16–18]. The chromophores that were taken into account in this study were the blood related absorbers deoxygenated-hemoglobin (Hb) and oxygenated-hemoglobin (HbO2), β-carotene, fat, water and collagen. The blood oxygen saturation (StO2) is defined as the fraction of hemoglobin that is oxygenated i.e. HbO2/(Hb + HbO2). The fat-to-water and fat fraction F/(W + F) is defined as fat content divided by the water and fat content of the tissue. For the scattering the empirical model described by
$$\upmu_{{\text{s}}{^{\prime}}} \left(\uplambda \right)\text{ = S800}\left[ {\text{Fmie}\left( {\frac{\uplambda}{{\uplambda_{\text{0}} }}} \right)^{ -{\text{ b}}} + \left( 1{ - } {\text{Fmie}} \right)\left( {\frac{\uplambda}{{\uplambda_{\text{0}} }}} \right)^{ - {\text{4}}} } \right]$$
(1)
is used where \(\uplambda_{0}\) = 800 nm corresponds to a wavelength normalization value, S800 is the reduced scattering amplitude at λ0, the Mie scattering slope is b, and Fmie denotes the Mie-to-total reduced scattering fraction assuming that Mie and Rayleigh scattering are the two dominant types of scattering in tissue.
Wavelength dependent features
Each fit parameter and principal component relates to the full wavelength range of the spectrum, so we also developed a method whereby each feature relates to a certain wavelength range. In this third method the spectrum was divided into equally-sized wavelength segments. For each segment a feature was calculated according to how the spectrum slopes in the segment as compared to the averaged spectra. To avoid the influence of intensity variations this was performed on normalized spectra. To determine the difference from the average spectra, the intensities were scaled for each wavelength, so that the average intensity was 0 and the standard deviation was 1. The scaled intensities were then added to obtain a scalar feature. Segment values are strongly positive when a spectrum slopes significantly more upward than the average, strongly negative when it slopes more downwards, and around zero when it behaves averagely or when the individual spectrum crosses the average spectrum.
This segment feature extraction method was considered as representative for the many other methods that can be used to describe the (relative) behavior of a curve.
Classification methods
Different classification methods were investigated to discriminate between fascicular tissue of the nerve and non-fascicular tissue (muscle, subcutaneous and sliding fat). These included: partial least squares discriminant analysis (PLS-DA), support vector machine (SVM) and classification and regression trees (CART) analysis.
Partial least squares discriminant analysis (PLS-DA)
Partial least-squares (PLS) [19] analysis is a regression method to find a linear relationship between a response variable Y (tissue type class) and the independent variable X (spectra). PLS-DA is widely used for analysis of spectra. The method is based on finding a number of principal components that represent as much of the variance in X as is possible. PLS-DA selects the principal components which are most relevant to the response variable Y. Therefore, PLS-DA acts directly on the spectra without the need for prior feature extraction. The PLS model is generated using a training data set. A discriminant analysis (DA) method is subsequently performed to obtain thresholds for discriminating the different responses (i.e. tissue classes). Prediction of class (tissue type) on the remaining data (the validation data set) is obtained by comparing the predicted PLS scores with the DA thresholds. The measured tissue type is assigned to one of the two predefined tissue classes depending on the PLS scores. The PLS-DA algorithm scripts used PLS Toolbox 6.2 (Eigenvector Research, Inc, Wenatchee, WA).
Support vector machine (SVM)
Support vector machines (SVM) [20] divide the feature space with a hyperplane that separates the two classes. In case full separation is not possible, a penalty parameter determines the behavior of the boundary. We used LIBSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) with a radial Kernel and standard parameters. Before using SVM, all features were scaled to a mean of 0 and a standard deviation of 1.
Classification and regression trees (CART) analysis
The classification and regression tree (CART) is based on a binary recursive partitioning algorithm. It starts from a central node that discriminates the two classes based on the best discriminating feature. From this node, daughter partial trees are generated and other features or the same are used for further splits. The purity of each node is assessed with the Gini’s maximization index algorithm which corresponds to unity minus the sum of square of the proportions of target classes at a specific node [21].
Statistical analysis
All classification methods used a training data set to create a classification model. A second, validation data set was then used to calculate accuracy, sensitivity and specificity for the model based on the confusion matrix of the validation data. Sensitivity and specificity are complementary (one can improve sensitivity at the cost of specificity and vice versa) so Matthews correlation coefficient (MCC) was used as a scalar measure for the quality of classification. MCC is given by [22]:
$$MCC = \frac{TP \times TN - FP \times FN}{{\sqrt {(TP + FP)(TP + FN)(TN + FP)(TN + FN)} }},$$
(2)
where TP are the true positives, TN the true negatives, FP the false positives and FN the false negatives. Here, fascicular tissue of the nerve was considered as a “positive” and surrounding non-fascicular tissues as a “negative”. The value of MCC is between −1 and 1, where 1 represents a perfect prediction, 0 no better than random and −1 total disagreement between prediction and actual state.
For the cervical nerve dataset a leave-one-specimen-out cross-validation approach was used, meaning the classification models were trained on the spectra from all but one cadaver and the validation was done on the spectra from the cadaver left out. This was repeated until all cadavers had been left out once, with the confusion matrices being added up.
The datasets acquired for the median nerve and surrounding tissue were classified by training the algorithms on the cervical nerve and surrounding tissues database only and then applying the model to the datasets collected in the area of the median nerve. This was to show how the classification algorithms work on truly independent data.