Collaboration:UNC-Utah-PNL 2006ProjectWk
From NAMIC Wiki
Home < Collaboration:UNC-Utah-PNL 2006ProjectWk
Contents
2006 Project Week: Meeting with UNC and Utah at the PNL | June 28, 2006
Present
Martha Shenton (PNL)
Marek Kubicki (PNL)
Marc Niethammer (PNL)
Sylvain Bouix (PNL)
Doug Markant (PNL)
Tom Fletcher (Utah)
Ross Whitaker (Utah)
Martin Styner (UNC)
Diffusion Tensor Imaging
- DWI measurements and construction of tensors
- Rician noise model (Utah): MRI has Rician noise (not Gaussian), though most models do not account for this distribution of noise, which ends at zero (no negative values) and has a long positive tail. This distribution produces a more positive mean, meaning a positive bias in measurements, thereby underestimating the amount of diffusion. This bias will be greatest in directions of high diffusion. Monte Carlo simulations demonstrated that Rician noise produces different answers based solely on whether tensors are aligned or misaligned with the gradient directions, and can even produce significant differences based on this variability.
- Increase in number of gradient directions helps to reduce the dependence on direction, but there will still be an overall bias
- Goal: To introduce tools into ITK that remove Rician noise
- Utah method combines spatial filtering with removal of Rician noise:
- level of Rician noise estimated from background signal
- best results found when filtering the DWIs
- tensor methods usually perform worse than DW methods
- used maximum likelihood estimation tool to test this on a real dataset
- Utah method combines spatial filtering with removal of Rician noise:
- Recommendations:
- Smoothing should be done in DW rather than tensor space
- Tensor estimation should be non-linear, though this is more expensive (40-60 minutes for a single volume); the PNL currently uses linear estimation
- This filtering would be an important application for fiber tractography at the PNL. PNL has tested using Gaussian smoothing on data to improve tractography; smoothing produces very different results with our data.
- Implementation:
- Validation study: PNL will take a previous dataset in which we found significant differences, measure again using Rician noise filtering. Expect to find greatest differences between studies in areas of high FA.
- Marek and Doug will figure out which ROI to use; Doug will upload data for Utah to test.
- Tractography: PNL will test results of filtering on tractography ( Marc and Tom )
- Validation study: PNL will take a previous dataset in which we found significant differences, measure again using Rician noise filtering. Expect to find greatest differences between studies in areas of high FA.
- Rician noise model (Utah): MRI has Rician noise (not Gaussian), though most models do not account for this distribution of noise, which ends at zero (no negative values) and has a long positive tail. This distribution produces a more positive mean, meaning a positive bias in measurements, thereby underestimating the amount of diffusion. This bias will be greatest in directions of high diffusion. Monte Carlo simulations demonstrated that Rician noise produces different answers based solely on whether tensors are aligned or misaligned with the gradient directions, and can even produce significant differences based on this variability.
- Previous tests between FA and GA
- Utah did not find differences from standard methods when using GA measure. At UNC, they found that GA is more sensitive to noise than FA, but they are still using FA since it is the standard in the field.
- GA measure is mathematically interesting (and correct), but there is no compelling reason to switch to this measure when doing analysis.
- Utah did not find differences from standard methods when using GA measure. At UNC, they found that GA is more sensitive to noise than FA, but they are still using FA since it is the standard in the field.
Shape Analysis
- PNL has been doing shape analysis on the caudate data on the NAMIC Wiki (N=30), as are other groups. All methods are being tested on the caudate, and converge globally in showing lateral and anterior shape differences (locally, there is divergence). The caudate data should be used by Utah/UNC as benchmark, starting with the smaller male sample.
- Results on this data in papers:
- James Levitt, CF Westin paper
- Bouix (MICCAI 2004) - solution to Poisson equations; collapsing surface to a sphere, measuring complexity
- Unpublished results from Martin using spherical harmonics
- Caudates were hand-segmented in Slicer, using rules for connecting topology (e.g., tail is cut off); Jim's paper has description of these manual measures, PNL can also describe these over email.* Utah has been testing different approaches to placing correspondences (parameterizations).
- Results on this data in papers:
- Some areas would be too difficult to study using spherical representations, for example cortex, STG, temporal lobe (variability between subjects is problematic for spherical harmonics correspondences).
- Ultimate goal would be to analyze the whole striatum. Jim Levitt (PNL) has manually parcellated the striatum, and is interested in regions of the striatum connecting to the frontal lobe.
