FreeSurfer

From SBIwiki

Jump to: navigation, search

Bias intensity field correction when using 32 ch array at 3T.

With 32 ch ( actually with 8 ch array as well) use the same pad thickness under the head all the time to keep the same distance between the occipital lobe and the coil!

This is an essential preprocessing step in T1w image segmentation. It is assumed a proper BET skull removal that alway precedes the bias field correction. N3 is a default bias correction in FS with 200 mm spline smoothing distance that works at 1.5 T and 8 ch. array but will fail at 3T and will certainly fail if 32 ch was used. This results in misclassification of tissue GM/WM in occipital, temporal and frontal lobes.

The accuracy and performance of N3 are affected by two key factors: firstly the choice of mask used to identify the region of the scan over which N3 works and secondly the estimate of the distance over which the non-uniformity field varies.

The sensitivity of the FreeSurfer segmentation approach to receiver array bias field relying on a mixture of edge based and statistical classification segmentation algorithms (Fischl et al., 2002, 2004b; Han and Fischl, 2007). Thus, while FreeSurfer's surface estimation is based on a deformable algorithm, the initialization of the surface deformation is provided by initial coarse segmentation, which is based on statistical classification and facilitated by a probabilistic atlas. An error in initialization can lead to the deformable surface algorithm being trapped at a wrong edge (Dale et al., 1999; Mcinerney and Terzopoulos, 1996). Moreover, the energy functional that guides the WM surface deformation process contains a term that penalizes the sum of intensity variances inside each tissue, which can also make the segmentation sensitive to intensity non-uniformity.


Recommendation recon-all script should be used with -expert file (to read-in expert options file, that must be edited before the use) and may be -3t option.

I case of troublesome brain region the that persistently fail in segmentation the coordinate file for these region should be supplied as argument (usually control.dat) [1] :

   'mri_normalize --f <path to file> '

Autorecon Processing Stages (default):

   1.  Motion Correction and Conform
   2.  NU (Non-Uniform intensity normalization N3) ---> options must be changed
   3.  Talairach transform computation
   4.  Intensity Normalization 1
   5.  Skull Strip ---> should be used as step 2, after 1.

Autorecon Processing Stages (rewised):

    2.   Skull Strip.  Restricting the algorithm to the intracranial cavity is ideal, as it should consist of three distinct tissue classes: CSF, GM and WM. 

Thus, a mask that correctly identifies the brain and CSF spaces should help N3 to improve the correction.

    3. NU (Non-Uniform intensity normalization N3) 
       Parameter modifications , which control smoothing distance and maximum number of iterations respectively are achieved by supplying two additional arguments to ‘mri_nu_correct.mni’ function:
       ‘mri_nu_correct.mni --distance 50 --proto-iters 500 ’


Options for N3: 1) the smoothing (spline) distance should be reduced from default value 200 mm to 50 mm.

The choice of smoothing distance is always a delicate compromise: the target is not to over-smooth the natural variation of GM-WM intensity in the image. The 50 mm is a safe but lower number like 30 mm (limit for N3) may overdo the job.

2) the maximum number of iterations from default 50 should be increased to 200- 1000 in order to cope with the smaller smoothing distances, which lead to longer convergence times and hence, require more iterations.

3) option --fwhm may also help



Ref.: Improvement of brain segmentation accuracy by optimizing non-uniformity correction using N3., Weili Zheng, NI, 2009

Example for segmentation with two spline smoothing distances:


File:FS N3 Zheng NI2009.png


Bias field estimation at several spline smoothing distances, demonstrate the over-smoothing:



File:FS-N3 Zheng NI2009 BF.png

Personal tools
MEG