Normalization Papers

Useful Papers - Normalization


Salmond et. al (2002), "The precision of anatomical normalization in the medial temporal lobe using spatial basis functions," NeuroImage 17(1), 507-12 PDF

Summary: The authors evaluate how closely they can get hand-drawn anatomical markers in MTL to line up as a function of their normalization parameters. They vary their number of nonlinear basis functions, their degree of regularization, and play around with templates a little bit. Results indicate that increasing constraints on the normalization actually increased their degree of anatomical precision.

Bottom line: A smaller number of nonlinear basis functions (4 x 5 x 4) is better for normalizing than a larger set (7 x 8 x 7). Increasing regularization (from medium to high) made the smaller basis set worse, but didn't improve the larger set's performance. Template choice (between a pediatric template and an adult) didn't make a lot of difference, even for pediatric data - but this is only in MTL...

Wilke et. al (2002), "Assessment of spatial normalization of whole-brain magnetic resonance images in children," Human Brain Mapping 17(1), 48-60 PDF

Summary: Using structural scans from 150 or so children of varying ages, Wilke et. al construct a pediatric template and then normalized their pediatric scans to both the new pediatric template and an adult template. They then examined the extent of deformation in each scan to find correlations with age and/or region as to how much deformation was needed to bring each scan into line with each template.

Bottom line: Lots of conclusions, but some main ones: brain size doesn't change much with age, but head size does, a lot; using the pediatric template required quite a bit less deformation overall, and there were several regions particularly affected by normalization to the adult template (prefrontal, temporal pole, etc.); even with the pediatric template, several regions are more variable than others in children (precuneus, prefrontal).


Crivello et. al (2002), "Comparison of spatial normalization procedures and their impact on functional maps," Human Brain Mapping 16(4), 228-50 PDF

Summary: The authors compared the performance of several normalization algorithms (affine, AIR's nonlinear, SPM, and a full-mesh method) in bringing segmented structural scans precisely in line with segmented templates, and on PET functional data.

Bottom line: The full-mesh method did best at bringing gray and white matter precisely into line with the templates, but the AIR nonlinear method and SPM both have the advantage of preserving local relationships between sulci and gyri more precisely. Functionally, although results differed between all four algorithms, none appeared to be significantly better than any others at relatively high PET resolution.

Ashburner & Friston (1999), "Nonlinear spatial normalization using basis functions," Human Brain Mapping 7(4), 254-66 PDF

Summary: Perhaps the classic paper about nonlinear normalization, and the defining paper for reference on how SPM does normalization. Also an excellent survey of the major issues about doing normalization, including their case for using voxel-based (as opposed to tensor-based) methods in SPM. Pretty technical, but highly recommended.

Bottom line: Any image can be transformed into any other image given enough time and an unconstrained linear transformation. But a precise transformation doesn't always get you what you want, which is overlap of functional areas. So the question is what constraints/priors will give you the best functional overlap and best performance.

Kochunov et. al (2000), "Evaluation of octree regional spatial normalization method for regional anatomical matching," Human Brain Mapping 11(3), 193-206 PDF

Summary: Can't say I got to this one, but it lays out a different method than the nonlinear-basis-function method that's standard these days.

Brett et. al (2001), "Spatial normalization of brain images with focal lesions using cost function masking," NeuroImage 14(2), 507-12 PDF

Summary: Can't say I got to this one either, but it's easily summarized: trying to line up images with big holes in them with templates is problematic when you allow nonlinear transformations, because the algorithm will tend to distort the lesion area and whole brain. An easy way to avoid that problem is by simply masking out the lesioned voxels when calculating your cost function, and Brett et. al demonstrate it works.