6.26 JHU DTI-based white-matter atlases

http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases

Two white-matter atlases provided by Dr.Susumu Mori, Laboratory of Brain Anatomical MRI, JHU.

• ICBM-DTI-81 white matter labels atlas
• 48 white matter tract labels
• hand segmentation
• standard-space average of diffusion MRI tensor maps
• 81 subjects
• mean age: 39
• JHU white-matter tractography atlas
• 20 structures
• identified probablistically
• average the deterministic tractography
• 28 normal subjects
• mean age: 29

Donna: When get to the TBSS stage, should talk about the WM atlas in more detail. Comparison of the pros/cons for each atlas.
MNI152 standard-space T1-weighted average structural template image (Smith, 2006)

• 152 structural images
• averaged after high-dimensional nonlinear registration into the common MNI152 co-ordinate system.
• 153 T1-volumetric brain MRI scans from the Oxford Project to Investigate Memory and Ageing (OPTIMA) were used. 1.5 Tesla dataset, spatial resolution: 0.86 * 0.86 * 2mm.

Read the reference: Grabner G, Janke AL, Budge MM, Smith D, Pruessner J, Collins DL (2006). Symmetric Atlasing and Model Based Segmentation: An Application to the Hippocampus in Older Adults. MICCAI; 4191: 58-66.
The eventual goal of all these methods is to produce an average model that can accurately both show the similarities in the group and exclude the effect of anatomic variance.
The technique presented here is different in that at each stage images are matched to the model in both their original and (Left-Right) mirrored orientation. The aim of this methodology were to:

1. Build a high resolution symmetric atlas from a population;
2. Automatically extract sub-cortical structures in a population;
3. Allow left-right volumetric comparisons to be made without bias;
4. Eliminate subjective error.

This paper has presented a method whereby automated model based segmentation of symmetric structures is made possible. Done by registering complementary original and flipped images to a target simultaneously, which is important to ensure bias is not introduced at any stage of the model generation. During iterations, registration begins with each of the images in their initial space, such constraint ensures that no bias is introduced towards a particular intermediate model during the construction of the final model.
This work embodies a small extension to the work done by Guimond on iterative atlasing but with two differences: 1) after each iteration of model building, the inverse of the average transformation is applied to the model, which controls for possible bias in the chosen method of registration. 2) in this case we are building a symmetric model.
Due to the nature and extent of processing required to build such a model, the computing time taken to generate this model is large. Approximately 2 weeks of dedicated processing time on a cluster of 16 Linus machines. But, once a model is built, the time taken to then match a novel subjects anatomy to the model is on the order of minutes.
One of the problems is the issue of symmetric bias when comparing structures left-right. The methodologies presented in this paper are intended to enhance the existing methods whereby models are built using non-linear registration with no corresponding impact on accuracy of automated segmentation. The most obvious benefit is the ability to perform comparisons of structures that are by nature assymmetric but posses enough symmetry to allow left-right comparisons to be made.

Dieter: has age been considered when the choosing between the two atlases?
(7.6) Dr.Joseph Schacht's response: the actual ROIs were the conjunction of the common white matter skeleton from TBSS and the JHU ROI because it was the best age match to my (younger) subjects and because it's nicely integrated with the FSL package. I do like that it's probablistic (I like the Harvard-Oxford gray matter atlas for the same reason), but this wasn't a big factor in my decision.