Chapter 9 Further analysis of the in vivo spectra was divided into two wavelength response ranges: Si‐sensor (≤1000 nm) and InGaAs‐sensor (≥900 nm). The number of extracted features (n=36) is too large to perform a statistically meaningful classification, as the extracted features could be redundant in the information they retain. Therefore, using combinations of all 36 features to build a classifier would result in a dimensionality problem and over‐fitting. We identified the most distinctive features, for classification of nerve in an adipose surrounding, by using binary logistic regression (SPSS Inc., Chicago, IL, USA) for both wavelength regions separately. This is a statistical technique that allows the prediction of categorical dependent variables (here the tissue type: nerve or adipose tissue) using a set of independent variables (here the spectral features)28. Using the approach for tissue classification based on hyperspectral data reported earlier by Akbari et al.29, we used support vector machine (SVM) to classify nerve within adipose surrounding. We used a polynomial kernel function30 for both wavelength regions. The SVM classifier attempts to find an optimum line in the two‐dimensional feature space, consisting of support vectors, to separate the training data with a minimum risk31. To estimate classification performance and to prevent overly optimistic results18‐20, we implemented a cross–validation (CV) approach. With a goal to obtain the classification accuracy as a performance measure, the data set is divided into a training set (to train the classifier) and a test set (to validate the classifier). First the CV approach uses leave‐one‐out (LOO) validation of nerve and adipose hyperspectral data acquired during thyroid and parathyroid surgery. This approach utilizes the same data set for both training and testing purposes and is very useful in cases of a relatively small data sample. Additionally the CV approach uses train‐test (TT) validation. In this approach we divided the data into a train set (consisting of nerve and adipose hyperspectral data acquired during thyroid and parathyroid surgery) and a test set (consisting of nerve and adipose hyperspectral data acquired during carpal tunnel release surgery). TT validation provides additional information as it estimates the performance of this system in the clinical setting by expanding the validation to other anatomical sites. Sensitivity, specificity, positive predictive value, negative predictive value and accuracy were calculated to quantify the classification performance for both wavelength regions (i.e. Si‐sensor and InGaAs‐sensor detection range), and for both cross‐validation methods (LOO and TT). In‐house developed classifiers (using MATLAB environment Version 7.7.0, MathWorks Inc., Natick, Massachussetts, USA) were used to estimate the classification performance. 134
proefschrift_Schols_SLV
To see the actual publication please follow the link above