Hi Warren
Wow, that is pretty bad. Sure, upload one and I'll take a look. Bruce
P.s. body transmit is probably a good thing. It's body receive you wouldn't want
On Jun 25, 2012, at 10:14 PM, "Winter, Warren" Warren.Winter@childrens.harvard.edu wrote:
Hi there,
We have a set of about 80 MPRAGEs whose acquisition parameters (chosen by and used at another lab with whom we collaborated) must have compromised their quality, as FreeSurfer has been consistently generating, in most of our subjects, a pial surface that excludes a lot of lateral cortex throughout the brain. The intensity values of true GM typically run from 40-80. The white matter surfaces are generally fine, though the intensity values drop off to the low 90s in the outer regions of WM strands. The skull strips are fine. We've had some success in using control points to push the pial surface out appropriately, but only after adding roughly between 300 and 1300 CPs per subject -- and even then it has not always fully ameliorated the problem.
Snapshots of a typical brain from this dataset: http://imgur.com/905Rm , http://imgur.com/5zVLQ , http://imgur.com/Abz2C . If it would help, I could upload a full subject to the file drop.
I ran mri_cnr on a bunch of our norm.mgz volumes from this data set and the gray/csf CNR is typically in the 0.55-0.80 range, whereas the gray/white CNR is around 1.7-2.0. To compare, bert's norm.mgz gray/csf CNR is ~1.05 CNR, gray/white CNR ~1.85. Here are the parameters that were used: TR = 1910 msec, TE = 3 msec, flip angle = 8 degrees. They are transverse T1-weighted MPRAGE images. Mri_probedicom output says that the transmitting coil was a body coil, which came as a surprise to us and is hopefully not true. In any event, these scans were of a pediatric population studied under special historical circumstances and so it is imperative that we salvage as much reliable data from the scans as possible.
I have tried a few other tweaks to recon-all in an attempt to automate some of the correction: (1) running as many as 8 iterations of nu_correct (default = 2) (2) running mri_normalize with b = 20 (from the wiki, this argues "use control point with intensity b below target (default=10.0)") and n = 5 (argues "use n 3d normalization iterations (default=2)") both separately and together (3) generating a bias field volume from 18 brains that had been successfully corrected with many control points (using mri_compute_bias), applying the bias volume to the orig.mgz (using mri_apply_bias) of other brains with the same issue, and then feeding this orig.mgz into recon-all with motion correction skipped and the first talaraich transform cleaned.
None of these adjustments has resulted in improvement comparable to massive control point placement. Is there anything else you could recommend? I've found in the archives some people reporting improvement for similar problems after using many more iterations of nu_correct than I tried, and I'm wondering if our images would benefit from more smoothing than is default (and if so, at which step -- as an argument in mri_compute_bias? As an expert option for mris_smooth?). Bruce also suggested in one case of low gray/csf contrast ( https://mail.nmr.mgh.harvard.edu/pipermail//freesurfer/2009-August/011572.ht... ) to manually correct an image and then use certain switches that will take the intensity distribution parameters from there rather than from the atlas.
Many thanks for your attention and help!
Warren
-- Warren Winter Research Coordinator Boston Children's Hospital Sheridan Laboratory of Cognitive Neuroscience Division of Developmental Medicine 1 Autumn Street, AU 650 Boston, MA 02215 857-218-5224
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer