Comparing classification methods for diffuse reflectance spectra to improve tissue specific laser surgery

Background: In the field of oral and maxillofacial surgery, newly developed laser scalpels have multiple advantages over traditional metal scalpels. However, they lack haptic feedback. This is dangerous near e. g. nerve tissue, which has to be preserved during surgery. One solution to this problem is to train an algorithm that analyzes the reflected light spectra during surgery and can classify these spectra into different tissue types, in order to ultimately send a warning or temporarily switch off the laser when critical tissue is about to be ablated. Various machine learning algorithms are available for this task, but a detailed analysis is needed to assess the most appropriate algorithm. Methods: In this study, a small data set is used to simulate many larger data sets according to a multivariate Gaussian distribution. Various machine learning algorithms are then trained and evaluated on these data sets. The algorithms' performance is subsequently evaluated and compared by averaged confusion matrices and ultimately by boxplots of misclassification rates. The results are validated on the smaller, experimental data set. Results: Most classifiers have a median misclassification rate below 0.25 in the simulated data. The most notable performance was observed for the Penalized Discriminant Analysis, with a misclassifiaction rate of 0.00 in the simulated data, and an average misclassification rate of 0.02 in a 10-fold cross validation on the original data. Conclusion: The results suggest a Penalized Discriminant Analysis is the most promising approach, most probably because it considers the functional, correlated nature of the reflectance spectra. The results of this study improve the accuracy of real-time tissue discrimination and are an essential step towards improving the safety of oral laser surgery.