🚀 go-pugleaf

Hi Lester,

Lester Ingber wrote:
> :CHEATS TO AVOID GLOBAL SEARCHES i.e. initial pose problem
> :---------------------------------------------------------
> 
> This of course is a very good idea -- to explore more efficient
> algorithms that might be better suited to your system.
> 
> :1) NOT a real solution. Constrain your image orientation with
> :   e.g. head holders. Also use axial images only, etc.
> 
> Will this work well across a large set of subjects?

This is THE METHOD OF CHOICE (because of its simplicity) for many
institutions around the world right now. 
Actually, it is an implicit (i.e. unmentioned) pre-requisite for all 
"automated" methods that I've come across so far except for that JNM
article of mine[*]. As I've mentioned in the last posting, all other
"automated" methods require the head orientation to be constrained to
within e.g. 30 deg. for these methods to work. i.e. these methods need
to know which direction "left" is, which direction "top" is, etc.

[*] Tsao J, Stundzia A & Ichise M. Fully automated detection of the 
    midsagittal plane for Tc-99m ECD SPECT brain images in 6 degrees of
    freedom. J Nucl Med 37(5): 8P. 1996. (abstract)

    The title actually understates the work because this method actually
    determines the entire brain orientation (i.e. the full coordinate
system),
    rather than just the mid-sagittal plane. More details will appear in
a
    recently submitted JNM paper.


> :2) Use centroid-based and principal axis transformation (i.e. higher-
> :   moment-based method to establish rough image orientation).
> :   PROBLEM: centroid-based method works only for estimating
> :            translation. Principal axis transformation works (NOT
> :            REALLY, see later) if the two shapes to be coregistered
> :            are similar. (e.g. Nat Alpert's work on brain registration)
> 
> I think such transformations run into trouble very quickly when dealing
> with nonlinear and/or noisy systems, especially in systems where the
> noise itself has nonlinear structure (likely true in brain imaging
> data).

The use of a convex hull adds stability to the calculations. (A convex
hull can be loosely defined as the smallest 3D convex shape that can
encase an object.) The reason that the convex hull adds stability to the
calculations is that even if there are slight shape variations due
to either noise or normal inter-subject variations, this variability is
often "masked" by the convex hull (i.e. these differences often don't 
show up in the convex hull.)

HOWEVER, the main problem with principal axes is that you have to be
matching similar objects. If you have a SPECT (Single Photon Emission
Computed Tomography) image which shows the amount of radioactive "dye"
in the brain circulation versus an MRI (magnetic resonance imaging)
image which shows anatomical structure of the head, you've got problems
because the SPECT can only visualize the brain (and not the scalp)
whereas the MRI visualizes ALL the structures (i.e. brain + surrounding
tissue + skull, etc.). Thus, the principal axes of these two images
will be completely different ---> Hence, this leads to the even more
troubling area of image segmentation: How can you cut away the scalp
from the brain when you DON'T even know how the brain is oriented.

CRANIAL SEGMENTATION:
If the scalp is very different in intensity from the brain, then
a simple thresholding operation will do. Otherwise, existing methods
require some "hacks" to combine morphological operations (e.g. dilution,
erosion, etc.) with thresholding to gradually remove the scalp.
You can't even do the segmentation by overlaying a model (a.k.a. 
a template) on your images to tell you where to segment because
*REMEMBER* you haven't figured how the brain is oriented in 3D
(6 deg. of freedom) yet!!! This in turns leads back to --->
Image registration in 6 deg. of freedom. DA DA!!! Quite frustrating,
really.

If anyone has some robust solutions to any part of this vicious cycle,
I would really like to hear about them.

Cordially,

Jeff
---------------------------------------------------------------------
Jeffrey Tsao, M.Sc.
Biomedical Magnetic Resonance Laboratory
University of Illinois at Urbana-Champaign