Recent Trends in MRI Brain-Tissue Segmentation
*†**†
*Department of Neurology, University of Minnesota, USA
†Department of Radiology, University of Minnesota, USA
Modeling & Analysis
Abstract
MRI image segmentation plays a key role in a variety of image processing applications including cortical surface extraction, the determination of cortical thickness and substructure volumes, intersubject and cross-modality registration, partial volume correction for radioisotope studies, and longitudinal studies of cerebral atrophy.
Numerous techniques with different data requirements and algorithmic assumptions have been employed for this purpose. In order to characterize the tissue segmentation algorithms used by the neuroimaging research community, we reviewed the recent biomedical literature and classified these algorithms along four dimensions of interest.
Methods
Peer-reviewed papers describing brain-tissue segmentation listed in Ovid Full Text for the period 1999.01.01 - 2002.12.31 were selected, and the segmentation algorithms classified according to: the number of input MRI volumes, e.g., T1, T2, proton-density (PD), diffusion-weighted (DW); degree of automaticity; output type ("hard" or "fuzzy" segmentation); and the amount of external information required (e.g., seed locations, template volumes, probabilistic atlases). Algorithms with mixed-tissue classes were treated as intermediate on the fuzziness scale, as were algorithms which produced bimodal tissue-fraction histograms.
Results/Discussion
There was no clear trend regarding the number of input volumes (typically T1, T1+T2, or T1+T2+PD). Although one might expect that the use of multiple input volumes would result in an increased ability to determine the fractional tissue content per voxel, there was no obvious correlation between the number of input volumes employed and the type of output produced, i.e., hard or fuzzy, and those methods which utilized more than three input volumes produced hard segmentations. We were surprised to discover that the majority of algorithms produced hard segmentations - in spite of the fact that, given the highly convoluted cerebral cortex with a thickness of 3-5 mm, significant volume averaging occurs even within isotropic 1 mm voxels. Although most algorithms employed unsupervised clustering or histogram thresholding techniques and did not require external inputs, the use of such inputs, e.g., the ICBM probabilistic atlas, to guide the segmentation appears to be increasing.
Acknowledgments
This work was supported in part by NIH grant MH57180.