Machine-Based Rejection of Low-Quality Spectra and Estimation of Brain Tumor Probabilities from Magnetic Resonance Spectroscopic Images

Magnetic resonance spectroscopic images (MRSI) carry spatially resolved information about the in vivo metabolism, however, their evaluation is difficult. Problems arise especially from artifacts and noise, yielding non-evaluable signals in many voxels. We propose a two-step approach to the processing of MRSI. In the first step a non-linear classifier is employed in every voxel to determine whether the spectral signal is evaluable, and if so, the tumor probability is computed in the second step. Thus, the quality control is strictly separated from the diagnostic evaluation of the spectrum. For an assessment of the proposed approach we consider MRSI-based brain tumor detection and localization and a tumor probability mapping by pattern recognition. In a quantitative comparison against the standard operator-controlled processing our interaction-free approach shows similar to superior performance.