To evaluate the quality of discrimination between the low-grade and high-grade lesions, linear discriminant analysis was applied. Specifically, we used the Karhunen-Loeve linear classifier, also known as the regularized linear discriminant function. For all spectroscopic modalities, the classifier was constructed using the leave-one-person-out approach, for which the classifier is constructed on a training set containing all spectra except the spectra measured in one particular patient. Compared to the leave-one-spectrum-out or leave-one-lesion-out methods, the leave-one-patient-out method improves the independence of the lesion classification since this approach prevents that spectra belonging to the same patient (which might not be considered to be completely independent) will be used both in the training data set (used to construct the classifier) and in the independent test set (used to test the goodness of the classification rule). Using the classifier, we obtain for all spectra a classification label, ie, low grade or high grade, and two posterior class probabilities (probability 1 and probability 2), which correspond to the confidence (from 0 to 1) of the classification rule about the membership of a certain spectrum to the low-grade or high-grade lesion classes. When multiple measurements were made on the same lesion, the final decision of the classifier for this lesion was obtained using the “mean fusion” approach. In the mean fusion approach, the posterior class probabilities obtained for all spectra measured on the same lesion were averaged, and the maximum value of these means of probability 1 and probability 2 were retained for the final decision of the classifier. The advantage of this approach is that the final decision on a lesion measured multiple times is an average between decisions made for each measurement separately.