Realignment Papers

Useful Papers - Realignment


Ardekani et. al (2001), "A quantitative comparison of motion detection algorithms," Magnetic Resonance Imaging 19, 959-963 PDF

Summary: Just what it says - tests SPM99, AFNI98, AIR, and TRU (pyramid) against each other on a data set generated from real data but with known misalignments. Evaluates how close each algorithm comes to correctness and how fast each runs.

Bottom line: SPM99 and AFNI98 outperform AIR and TRU by a wide margin, particularly with a big misalignment. AFNI98 is almost as good as SPM99 (better at some SNRs) and is wayyyy faster (we needed an experiment to tell us this?).

Bullmore et. al (1999), "Methods for diagnosis and treatment of stimulus-correlated motion in generic brain activation studies using fMRI," Human Brain Mapping 7, 38-48 PDF

Summary: Introduces the issue of task-correlated motion, demonstrates a between-group study that is severely biased by it, and introduces some methods to account for the degree of correlation and, in some small ways, correct for it.

Bottom line: Task-correlated motion can severely bias between-group studies, but you can compensate to some degree for it if you can measure the degree of the correlation.

E-mail thread (1999) between Field and Ashburner about entering motion parameters into your design matrix (click 'Next in Topic' once you're there to follow the thread along) Thread

E-mail thread (2000) between Flaisch, Henson, and others, more about motion parameters in the design (click 'Next in Topic' once you're there to follow the thread along) Thread

Summary: For both of these: Including your motion parameters in your design matrix is like regressing out motion-correlated signal, which can reduce your false positives at the expense of reducing your true positives. Values derived from the parameters (sines, squares, etc.) can also be useful.

Bottom line: Including these depends heavily on your own study - how much motion, how big your signal is. Some objective tests (like Skudlarski et. al PDF have shown, in general, that it doesn't make a huge difference either way. Probably worth testing on your own studies...


Field et. al (2000), "False cerebral activation on BOLD functional MR images: study of low-amplitude motion weakly correlated to stimulus," American Journal of Neuroradiology 21, 1388-1396 PDF

Summary: Uses a computer-controlled physical phantom to see if highly task-correlated but small motions can induce false activations in data.

Bottom line: Yes, in a big way. Even with submillimeter motions, head motion that correlated better than r = 0.52 or so routinely generated false activations - sometimes very realistic-looking and significant false activations. Evaluate the task correlation of your motion parameters!

Ward et. al (2001), "Prospective multiaxial motion correction for fMRI," Magnetic Resonance in Medicine 43, 459-469 PDF

Summary: Demonstrates the effectiveness of a prospective (i.e., during image acquisition) motion correction algorithm that handles 3D correction and removes worries about intensity changes or realignment validity.

Bottom line: This is one direction realignment research is going; early testing showed their algorithm effectively compensated for motion of up to 10mm and up to 10 degrees, but at the cost of 320 msec per TR.

Grootonk et. al (2000), "Characterization and correction of interpolation effects in the realignment of fMRI time series," NeuroImage 11, 49-57 PDF

Summary: Argues that residual motion-correlated intensity changes after realignment are largely the result of interpolation errors in the realignment, and that including the sines and cosines of your motion parameters in your design matrix can account for these.

Bottom line: The algorithm handles interpolation errors very well, but is still subject to the same concerns about including motion parameters in your design matrix discussed above...

Freire & Mangin (2001), "Motion correction algorithms may create spurious brain activations in the absence of subject motion," NeuroImage 14, 709-722 PDF

Summary: Introduces the problem with least-squares algorithms, and demonstrates how they can induce false activations with both simulated and real data. Also demonstrates that non-least-squares methods don't all suffer from the problem.

Bottom line: Motion parameters can be biased by activations with large signal changes; other algorithms like mutual information don't suffer from that problem.

Morgan et. al (2001), "Comparison of functional MRI image realignment tools using a computer-generated phantom," Magnetic Resonance in Medicine 46, 510-514 PDF

Summary: Similar to Ardekani et. al, but explicitly compares tools in terms of activation results, not just correctness of algorithm. Uses computer-generated data to compare SPM99, AFNI98, SPM96, AFNI96, and two types of AIR.

Bottom line: Most of the algorithms perform about the same for true-positive data (besides SPM96, which suffers), but SPM99 is better than the other at removing false positives.


E-mail thread (2003) from Jesper Andersson, who wrote the unwarping code, and others, about why unwarping is useful and how it differs from including your motion parameters in your matrix. (click 'Next in Topic' once you're there to follow the thread along) Thread

Summary: Good overview of unwarping from the guy that wrote it.

Bottom line: If you use EPI, it can be useful. Unwarping is a more highly targeted version of including your motion parameters in your design matrix - instead of taking out ALL motion-correlated variance (including real activations), it only knocks out motion-by-susceptibility artifacts, and those hopefully account for a good chunk of your motion artifact, particularly in high-susceptibility regions.

Andersson et. al (2001), "Modeling geometric distortions in EPI time series," NeuroImage 13, 903-919 PDF

Summary: Haven't read it yet meself, but this lays out the mathematical background for unwarping - describes the model of susceptibility-by-motion interactions that SPM now uses to calculate the deformation fields and remove their effects. They describe a further paper, but I don't know of it...

Birn et. al (1999), "Event-related fMRI of tasks involving brief motion," Human Brain Mapping 7, 106-115 PDF

Summary: Several tasks of potential psychological interest - talking, swallowing, tongue movement - involve head motion that introduces significant artifacts into fMRI images, making it difficult to distinguish task-related activation in these trials from artifactual activation. But because the temporal profiles of the two sources are different, it possible in principle to separate the two in event-related designs, which is what Birn et. al attempt here, with some success.