Validation of a method for coregistering scalp recording locations with 3D structural MR images

Abstract

A common problem in brain imaging is how to most appropriately coregister anatomical and functional data sets into a common space. For surface-based recordings such as the event related optical signal (EROS), near-infrared spectroscopy (NIRS), event-related potentials (ERPs), and magnetoencephalography (MEG), alignment is typically done using either (1) a landmark-based method involving placement of surface markers that can be detected in both modalities; or (2) surface-fitting alignment that samples many points on the surface of the head in the functional space and aligns those points to the surface of the anatomical image. Here we compare these two approaches and advocate a combination of the two in order to optimize coregistration of EROS and NIRS data with structural magnetic resonance images (sMRI). Digitized 3D sensor locations obtained with a Polhemus digitizer can be effectively coregistered with sMRI using fiducial alignment as an initial guess followed by a Marquardt-Levenberg least-squares rigid-body transform (df = 6) to match the surfaces. Additional scaling parameters (df = 3) and point-by-point surface constraints can also be employed to further improve fitting. These alignment procedures place the lower-bound residual error at 1.3 +/- 0.1 mm (micro +/- s) and the upper-bound target registration error at 4.4 +/- 0.6 mm (micro +/- s). The dependence of such errors on scalp segmentation, number of registration points, and initial guess is also investigated. By optimizing alignment techniques, anatomical localization of surface recordings can be improved in individual subjects.

Structure‐function relationships in the visual system in multiple sclerosis: an MEG and OCT study

“…The head‐localization coil positions were digitized, as well as the outline of the participant's scalp (~500 points), using a 3D digitizer (Fasttrak, Polhemus, Colchester, VT, USA). Scalp surfaces of all subjects were coregistered with their structural MRIs using a surface‐matching procedure, with an accuracy of about 4 mm 22. A single sphere was fitted to the outline of the scalp as obtained from the coregistered MRI, which was used as a volume conductor model for the beamformer approach described below.…”

Comparison of procedures for co-registering scalp-recording locations to anatomical magnetic resonance images

“…Therefore, the choice of the initial guess is very important. Consistent with this logic, Whalen et al 23 showed that LMA performance is strongly dependent on the accuracy of the initial guess. In other words, LMA worked well only when coupled with an initial guess based on fiducial alignment, and its performance degraded substantially when the initial guess was wrong.…”

“…Surface-fitting procedures [21][22][23][24] offer alternative or supplemental methods to fiducial-based approaches. They fit a discrete sampling of two surfaces, one represented by the digitized points and another represented by points extracted from a rendition of the scalp derived from an sMRI image.…”

“…The Levenberg and Marquardt algorithm (LMA) 25,26 also defines the objective function using Euclidean distance and the least-square method (or a modification of this method called damped least squares), but uses special procedures to explore the parameter space to reduce the probability of getting stuck in a local minimum. Whalen et al 23 showed that using fiducial alignment to generate the initial guess, followed by application of the LMA algorithm, outperforms fiducial alignment alone when co-registering the digitized points to the extracted head surface. For brain imaging scalp co-registration, the problem of local minima is particularly crucial because the surface of the head is nearly spherical/ellipsoidal in shape, and symmetries can easily cause erronous solutions.…”

“…In fact, a completely automated procedure was proposed by Kozinska et al 22 This procedure uses geometrical moments to generate the initial guess, but lacks an effective strategy, such as LMA, for exploring the parameter space and optimizing the co-registration. In this paper, we introduce an iterative co-registration method that, to our knowledge, was never used for this kind of application, ICP2P, 27,28 and we compare it to the pairing of fiducial alignment with LMA proposed by Whalen et al 23 A difference between ICP2P and LMA is that ICP2P relies on an objective function, based not only on Euclidean distances between the points of the two surfaces, but also on estimated normals (hence curvatures) of the destination surface. ICP2P emphasizes regions with high curvature for improving co-registration during the iterative procedure.…”