Brain lesion classification using 3T MRS spectra and paired SVM kernels

The increased power and resolution capabilities of 3T Magnetic Resonance (MR) scanners have extended the reach of Magnetic Resonance Spectroscopy as a non-invasive diagnostic tool. Practical sensor calibration issues, magnetic field homogeneity effects and measurement noise introduce distortion into the obtained spectra. Therefore, a combination of robust preprocessing models and nonlinear pattern analysis algorithms is needed in order to evaluate and map the underlying relations of the measured metabolites. The aim of this work is threefold. Firstly we propose the use of a paired support vector machine kernel utilizing metabolic data from both affected and normal voxels in the patient's brain for lesion classification problem. Secondly we quantify the performance of an optimal reduced feature set based on targeted CSI-144 scans in order to further reduce the data volume required for a reliable computed aided diagnosis. Thirdly we expand our previous formulation to full multiclass classification. The long term aim remains to provide the human expert with an easily interpretable system to assist clinicians with the time, volume and accuracy demanding diagnostic process.