A novel tumor grading technique using functional magnetic resonance imaging

Adequate therapy planning of gliomas needs histological determination of invasive biopsy due to the fact that both treatment and prognosis of glial neoplasms vary strongly depending on their histological grading. Magnetic resonance imaging based glioma grading is currently mainly based on contrast-enhanced T1-weighted images. To additionally gain information on tumor physiology for glioma grading functional MRI techniques like perfusion MR have also been considered. Here, we present a novel technique for glioma grading, which is based on a similarity search for uncertain data. In order to perform the search as accurate as possible we used four different features of the tumors as input for the similarity search: the Cerebral Blood Volume, the Cerebral Blood Flow, the Mean Transit Time, and the post-contrast T1-weighted image. For each patient a tumor feature vector was defined by a four-dimensional Probability Density Function, more precisely a Gaussian Mixture Model (GMM). In contrast to existing similarity searches we also considered correlations between different features in the similarity search. Here, we present the entire workflow from image acquisition to tumor grade prediction providing a probability-based classifier for gliomas. We applied our approach to MRI data sets of glioma patients, which were preprocessed and converted to four-dimensional GMMs, and achieved an accuracy, sensitivity, and specificity of 83.8%, 78.6%, and 87.0% while grading based solely on contrast-enhancement could only achieve an accuracy, sensitivity, and specificity of 64.9%, 52.4%, and 81.3%, respectively. Hence, our proposed similarity search based grading technique seems to be of great value for supporting non-invasive tumor grading since it integrates the information of different MRI sequences including perfusion maps in one semi-automatic analysis.