Increasing the credibility of MR spectroscopy-based automatic brain tumor classification systems

In the last decade many approaches have been introduced that allow for automatic classification of brain tumors by means of pattern recognition and magnetic resonance spectroscopy. Despite promising classification accuracies, none of these methods has found its way into clinical practice, which is also related to the missing transparency for the basis of their decision making. In this work, we develop two methods to increase the interpretability of such classification systems. First we propose a new reliability measure that determines a lower bound for the probability that a particular classification is correct. Additionally, we present a method that visualizes important regions for the classifier directly in the spectral domain. As a basis for this, seven classification methods were evaluated for their performance in discriminating aggressive tumors, low-grade glioma and meningioma, based on a common database. Our results show that the novel reliability measure is in good agreement with the actual classification accuracy. Further we point out that our visualization method clearly indicates which spectral regions are important for a classifier and how metabolite concentrations correspond to specific tumor types. Combining both methods can help to better understand a classifier's decision and therefore make the outcome more transparent and trustworthy.