Segmentation of proton MRS signals compromised by partial volume contamination.

Abstract

A method is presented for segmenting in vivo magnetic resonance spectroscopy (MRS) signals compromised by partial volume contamination (PVC). This method models PVC as a linear phenomenon and can be used with spectra acquired at any spatial resolution. It is designed to segment multivoxel or chemical shift imaging (CSI) data. The estimator used, ASPECT, inputs pristine and compromised signals and outputs fractional components of the pristine signals in the compromised signal. The method is evaluated using simulations and $\sp1$H-MRS data taken from phantoms and human brain. ASPECT is a sensitive estimator in simulations satisfying linearity; however, spectral perturbants arising from magnetic field inhomogeneity ($\Delta$B$\sb0$), eddy currents, and biological heterogeneity violate linearity and are sources of error. $\Delta$B$\sb0$ and eddy currents cause intervoxel frequency shifts and phase differences while biological heterogeneity influences signal quality. $\Delta$B$\sb0$ also causes spatially variant intravoxel dephasing and a concomitant loss in resonance intensity. Minimizing these errors is insufficient to satisfy linearity and corrections are introduced to reduce residual error in single voxel (SV) and CSI data. One correction eliminates eddy current induced spectral artifacts and produces pure absorption line shapes. In CSI data, it also eliminates intervoxel frequency shifts. In vitro evaluation shows that ASPECT accounts for the signal present when using data corrected for intervoxel frequency shifts and phase differences; however, spatially variant intravoxel dephasing produces error not accounted for. The magnitude of this error is largest with SV data and reduced with higher resolution CSI data retrospectively summed together to sample $\Delta$B$\sb0$ over the same volume. This is because $\Delta$B$\sb0$ is smaller over the CSI voxels; summing corrected CSI data eliminates intervoxel frequency shifts thus recovering some of the lost intensity. In vivo evaluation shows that ASPECT accounts for the majority of the signal present for corrected data; however, spatially variant intravoxel dephasing and some degree of PVC in the pristine signals produced nonlinearities not accounted for. This is a problem ASPECT shares with other estimators and suggests that their potential in vivo application is limited by the availability of pristine signal components.