| HBM 2005 Abstracts: Previous | Next | Click here for a PDF of this poster. |
1Department of Radiology, University of Minnesota, USA, 2Department of Psychology, University of Minnesota, USA, 3Department of Neurology, University of Minnesota, USA |
Objective:
Segmentation of MRI brain volumes into anatomical compartments, e.g., left cerebral hemisphere (LCH), right cerebral hemisphere (RCH), cerebellum (CB) and brain stem (B), is important for many biomedical and neuroscience applications. We here present a fully automatic method based on registration to a template volume that has a corresponding spatial-compartment mask. "Contested" spatial compartment boundaries are then transferred to the subject volume and refined.
Methods: A template dataset is prepared from a T1-weighted MRI brain volume. That volume is stripped (non-brain tissue excluded), and a spatial-compartment mask (LCH, RCH, CB, B) is prepared (in our case, manually). The subject volume must also be stripped -- but the stripping need not be perfect. However, the method can not distinguish brain tissue from blood vessels and will assign such voxels to their neighboring compartment. In stage 1, a fourth-order polynomial warp is computed using AIR [1] to align the template MRI volume to the subject volume. The warp transform is then independently applied to each of the template volume's four spatial compartments using trilinear interpolation. Subject voxels "claimed" by two or more compartments are designated "contested" (see Figure 1, upper row). In stage 2, the initial spatial compartments are refined. For each coronal slice a perimeter is computed for each spatial compartment from the clean boundary and the skeleton of any contested boundary. Subsequently the boundaries undergo "wiggling" within their local neighborhoods to settle into positions of low image intensity -- presumed to be CSF-weighted (See Figure 1, lower row). On completion of this stage global operations ensure that all voxels in the subject-volume mask receive a label and that each mask compartment is a single simply-connected region.
Results & Discussion:
Forty-five 3T SPGR scans (.9375 x .9375 x1.5 mm voxel size) of healthy subjects have been segmented using this approach. Figure 2 illustrates the left/right hemisphere divisions for four SPGR volumes. Repeat scans from two normal subjects (one scan each at 1.5T and 4T, each with 1 mm3 voxels) were also segmented (Figure 3). The upper right panel in Figure 3 exhibits minor "leakage" of the right hemisphere into the left occipital lobe. In each case compartment definitions were qualitatively satisfactory (aside from the occasional minor leakage), although anatomical boundaries are delimited by jagged lines. Run time per volume was on the order of 1 hour on a 3GHz processor.
Conclusions:
The refinement of 2D slice boundaries rather than 3D volume boundaries leads to compartment boundaries that are seldom smooth in all slice views. Although our method relies heavily on the quality of the registration and is computationally expensive, it produces consistent results across MRI volumes from different sites, subjects and scanning protocols.
References & Acknowledgements: |
|
|
|
