Chapter 9 Multimodality image coregistration for MRI-negative epilepsy surgery
MRI-Negative Epilepsy, ed. Elson L. So and Philippe Ryvlin. Published by Cambridge University Press. © Cambridge University Press 2015.
Introduction
Resective surgery in MRI-negative epilepsy results in poorer outcomes and a higher rate of recurrence compared to cases with a structural lesion visible on magnetic resonance imaging (MRI). As many as 26% overall and 46% of extratemporal (1) patients undergoing epilepsy surgery have negative MRI, and in these cases functional studies, including positron emission tomography (PET), ictal and interictal single photon emission computed tomography (SPECT), diffusion-weighted MRI (DWI), functional MRI (fMRI), and chronic intracranial EEG (icEEG) monitoring are often essential for localization of the epileptogenic zone. In order to coalesce the results from these disparate modalities around a coherent epileptogenic hypothesis to guide resective surgery, the functional data must be spatially aligned into a single coordinate system, typically corresponding to the patient’s high-resolution T1-weighted MRI, to ultimately guide the resection plan.
In addition, improvements to the sensitivity and specificity of functional modalities are often achievable by comparison of patient scans to spatially varying statistical metrics of normality derived from measurements of normal, or nonepileptic, subjects. For a group analysis, it is necessary to transform all data to one common template space. Recent results in cerebral blood flow mapping with ictal–interictal subtraction SPECT show significant improvement in localization when patient scans are evaluated in the context of paired resting scans of normal individuals (2,3). Statistical parametric mapping of [18F] fluorodeoxyglucose (FDG) PET scans in comparison to normals has been shown to improve the localizing capability of PET in epilepsy (4,5). The development of novel receptor PET tracers has also furthered the study of epilepsy via PET, and SPM will likely prove useful as normal scans become available.
Preprocessing for coregistration
Image coregistration is predicated upon the assumption that voxel intensities or features in one image correspond directly to intensities or features in the image to be coregistered. In some cases, especially cross-modality registration, that assumption requires images to be preprocessed before registration can proceed.
For accurate registration of functional imaging modalities to high-resolution MRI it is often beneficial to separate brain voxels from the entire head image. Elimination of nonbrain signals such as skull, orbits, and muscles generally improves the robustness of the registration, as functional images usually contain little extra-brain tissue, and structural images may contain changes in extracerebral morphology (e.g., a craniotomy in CT images of electrode placement). There are three commonly used methods for achieving brain/nonbrain segmentation: manual tracing, intensity thresholding with morphological processing, and model-based approach. Manual thresholding is arduous and time consuming, but does offer the user the opportunity to adjust appropriately to anomalies (tumor, prior resection, surgical intervention, etc.) during the segmentation process. Semi-automated intensity thresholding with morphological processing is much less arduous and may allow the user to retain some control over characterization of anatomical anomalies. Model-based cerebral segmentation typically nonlinearly maps an anatomical model to the brain either using surface or voxel-based criteria. The procedure requires very little user intervention and offers the added benefit of providing standardized atlas labels and regions of interest to the brain volume, but the procedure may inappropriately label anatomical features that do not correspond clearly to standard brain regions. For functional imaging modalities such as PET and SPECT, simple thresholding is usually sufficient to remove extracerebral activity.
Some intensity normalization or filtering may be necessary as well to facilitate image coregistration. The PET and SPECT images suffer from attenuation of emitted photons passing through brain and skull tissue and commensurate reduction in counting statistics proportional to depth from the skin surface. Most SPECT and PET acquisition systems apply attenuation correction as part of the image reconstruction process, and while the typical user would not need to correct for this effect explicitly, it does result in reduction of signal-to-noise ratio near the center of the brain. The MRI images may suffer from intensity nonuniformities due to uneven receiver coil coverage or an inhomogeneous main magnetic field, resulting in a bias field across the image. A number of techniques have been proposed for correcting this class of artifact (6,7), including filtering methods, histogram correction, and template-based methods. Other types of artifacts in the image due to patient motion, magnetic susceptibility, fluid flow, and other factors may distort the relationships between voxel intensities or features to be coregistered, and they would be expected to reduce the accuracy of image registration.
Image registration algorithms
The goodness of the match is based on a cost function (Figure 9.1), which is calculated using some optimization algorithm. The cost function may involve relationships between features (points, lines, surfaces, or anatomical features) in the images or a mathematical computation involving intensity values of corresponding image voxels. The cost function is designed such that when the two image sets are in perfect alignment the cost calculation reaches an unambiguous globally minimal1 value. Normally, there are many parameters, and it is not feasible to search through the whole parameter space. The usual approach is to make an initial parameter estimate, and begin iteratively searching from there. A judgment is then made about how the parameter estimates should be modified, before continuing on to the next iteration. The optimization is terminated when some convergence criterion is achieved (usually when chi2 stops decreasing or a fixed number of iterations is reached).
Figure 9.1 Normalized mutual information rigid body registration of a PET image to a T1-weighted MRI. The joint grayscale probability distributions are shown before and after registration.
Numerical minimization is the process by which the parameter space of rotations and translations is searched efficiently in order to find the optimal value of the registration cost function. A number of published strategies exist for cost function minimization, and the reader is directed to the literature for further detail (8,9). To speed up the optimization process and to avoid local minima, most currently used registration methods employ some form of multiresolution optimization. The images at larger scales are subsampled versions (often with preblurring) of the original high-resolution images, and so contain fewer voxels, which means that evaluating the cost function requires less computation. In addition, as only gross features of the images remain at these large scales, it is hoped that there will be fewer local minima for the optimization to be trapped in.
Interpolation, though often not considered part of the image registration algorithm itself, is an important step in the coregistration process and can be viewed as reconstructing a full continuous image from the set of discrete transformed points. Interpolation methods can significantly affect the quality of the registered image, and some precision has to be sacrificed in order to lower computational efforts. Nearest neighbor (zero-order hold resampling) provides very fast interpolation that maps to the value of the closest neighboring voxel without correcting for intensities at subvoxel displacement. This algorithm is very fast, but the resulting image quality is degraded quite considerably resulting in the resampled image having a blocky appearance. Bi- and trilinear interpolation (first-order hold) is slower than nearest neighbor, but the resulting images are of higher quality. These interpolation methods provide visually appealing output images, but do suffer from loss of high-frequency information due to averaging of neighboring pixel values in the interpolation process. Higher-degree interpolation algorithms (polynomial interpolation), including quadratic, cubic, spline, and Gaussian, provide more precise interpolation but are computationally slower because of more neighboring voxels used in the algorithm. Windowed sinc (sin(x)/x) interpolation can be used to apply image transformations and may offer the nearest approximation to true Fourier interpolation given appropriate choice of window (10).
Linear registration
Linear registration is an important component of medical image analysis as it enables in a single patient comparison of different imaging sequences, or multiple imaging modalities that can provide additional information not available in a single image modality. With intermodal coregistration, one image (usually a high-resolution structural MRI scan) remains stationary and other images are spatially transformed to match it. Linear rigid body registration is used routinely to correct movement between sequential scans in an acquisition, as some degree of subject motion within the scanner is usually present, especially when the scanning takes a long time. Once images are in the same space they can be averaged in order to increase signal to noise, or subtracted to emphasize differences between the images. Since the shape of the human brain changes very little with head movement, rigid body transformation can be used to correct for different head positions of the same subject. This is an important task as images from different imaging modalities can have different orientation, field of view size, contrast, and noise distribution.
Linear transformations are the most basic class of transformations, where all voxels are constrained to move according to a global, linear relationship. A rigid body transformation in three dimensions is defined by six parameters: three translations and three rotations. Scale terms are computed from the voxel dimensions of the images to account for global scale differences between images. Affine registration is similar to rigid body registration but scale and linear shear terms are calculated from the registration metric. While still a linear operation, affine registration can be thought of as a low-order, limited version of nonlinear registration or spatial normalization.
Feature and surface metrics
Feature-based techniques identify homologous features (labels) in the source and reference images and find the transformation for the best match. The features can be homologous points, lines, or surfaces. Corresponding features can be identified manually, but this makes this process subjective and time demanding. Anatomical landmark-based matching of corresponding points is the simplest form of registration. The user chooses homologous points in the base and transforming images, and a least-squares error minimization is performed to compute the best transformation between the sets of points. This procedure is user intensive, and accuracy can be poor unless a large number of homologous points are used, due to the imprecision of landmark identification (11), but it may be necessary in cases where homologous surfaces or voxel intensity correlations cannot be established. However, point-based registration with invasively fixed fiducial markers is used in neurosurgical systems with exceptional accuracy, and represents the “gold standard” for accurate image-guided surgical navigation (12).
Surface-based registration has been commonly used with good results. Surface registration algorithms require the user to identify analogous surfaces in the reference and transformation images, and a best fit between surfaces is calculated. The brain surface is typically used, although this may prove unreliable if significant shift has occurred with a craniotomy (13,14). Surface matching algorithms minimize the mean squared distance or some variant thereof between homologous surfaces in the two images, producing a “hat on head” best fit (15,16). Surface-based registrations lend themselves particularly well to applications where one of the data sets to be registered is inherently a surface, for example matching scalp EEG electrode positions to a subject’s MRI (17,18), or laser surface scanning for intraoperative image-guided navigation (19). However, for most high-resolution volumetric image registration applications, voxel-based cost functions have been shown to provide more accurate results than surface-based algorithms (20).
Voxel intensity metrics
Voxel-based algorithms assume certain relationship between voxel values between two images. This relationship can be a simple correspondence in the same modality registration (e.g., SPECT to SPECT) or a more complex relationship when one tissue voxel value corresponds to two or more tissues in other modality (e.g., CT to PET).
Woods et al. (21) developed one of the earliest successful voxel intensity-based algorithms for image registration. The algorithm’s cost measure of registration between MR and PET is based on the assumption that when registered the range of PET values associated with a particular value of MR should be minimized. The overall measure is a sum of the standard deviations of the ratios of corresponding image values, with an iterative Newton–Raphson method employed to minimize the cost function. While the algorithm has been used successfully in a variety of applications (22,23), there is a concern that Woods’ technique can break down when there is a bimodal or multimodal distribution of test volume values (24), which is often the case when matching CT and MRI, where CT voxel intensities of brain parenchyma show little grayscale variation.
Mutual information (24,25), and normalized mutual information (26,27), is a measure of dependence of one image on the other, and can be considered as the distance between the joint grayscale probability distribution (P(f; g)) and the grayscale probability distributions assuming complete independence (P(f)P(g)). When the two distributions are identical, this distance (and the mutual information) is zero. Mutual information-based registration does not assume a specific intensity relationship between different images and is a measure of statistical dependency between two data sets. The most widely adopted scheme for maximizing mutual information is Powell’s method, which involves a series of successive line searches for each dimension until convergence is reached. Mutual information registration can handle data that are conditionally multimodal.
While Woods’ algorithm and mutual information represent the most widely used voxel intensity-based image registration techniques, other cost functions are also used, and they include entropy correlation coefficient, and normalized cross-correlation (28). Cost function and interpolation are usually shipped in a registration package. These include AIR (21), SPM (29), UMDS (27), MRITOTAL (30), and FSL (31).
Nonlinear spatial normalization (warping)
Nonlinear spatial normalization involves transforming the subjects’ data into a normalized stereotactic space by a given matching criterion. Definition of normalized space is the same for all imaging modalities, so no further coregistration is needed once images are normalized. Spatial normalization may be achieved using many of the same cost functions underlying rigid body registration, but with higher-order polynomials or spatial frequency components modeling the transformation. A common reference frame enables data comparisons across subjects and time. The first step in registering images from different subjects involves determining the optimum 12 parameter affine transformation used to account for gross differences in head position and shape.
Lower-order scaling and skew terms
The rationale for adopting a low-dimensional approach is that it allows rapid and general modeling of global brain shape. With more registration parameters used, a more precise fit is usually achieved, but this relationship is not linear. Depending on the analysis design, most of the variability can be accounted for using only nine or 12 degrees of freedom, eliminating major differences in brain shape. With the exception of major sulci, there is a high interindividual variance in cortical shape and features; and to achieve a perfect match, it would be necessary to create sulci and gyri that do not exist in the original brain. There is also no guarantee that spatially normalized gray matter loci are functionally equivalent, particularly in patients with chronic epilepsy (32). Spatial smoothing, commonly used in spatial normalization, improves the signal-to-noise ratio and is able to compensate for small anatomical and functional interindividual variability with the cost of losing some high-frequency information.
Higher-order complex deformations
High-dimensional registration can involve thousands to millions of match parameters, and so is potentially able to match images very precisely. Additional registration parameters used require more computing power; so for speed, a relatively low number of parameters is usually used initially, with more complex parameters added progressively. Generally, the algorithms work by minimizing the sum of squared difference between the image which is to be normalized and a linear combination of one or more template images. For optimal estimation, contrast in the template and registered image should be similar.
The set of sulci that are consistently present in normal subject is quite limited, and it includes the interhemispheric fissure, the Sylvian fissure, the parietal–occipital fissure, and the central sulcus. The differences in sulcal pattern appear even in monozygotic twins (33). Additional functional and cytoarchitectonic differences can make definition of homologous areas problematic (34).
Visually, the registered images appear very similar following spatial normalization. This is not an adequate indication of the quality of the registration, but it does confirm that the optimization algorithm has reduced the likelihood potentials. It is possible for the wrong structures to be registered together, but distorted so that they look identical (35). Validation of warping methods is a complex area. Klein (36) evaluated 14 registration algorithms based on overlap measures of manually labeled anatomical regions and found only a modest correlation between the number of degrees of freedom of the deformation and registration accuracy.
For population studies, mapping to a standardized template space is often beneficial because it enables reporting various activations or brain changes in reference coordinates. These changes can then be compared across different studies. Furthermore, for a group analysis, it is necessary to have all brains in the same space. This template stereotactic space is usually defined by ICBM or MNI brain template and an approximate of the space described in the atlas of Talairach and Tournoux (1988). An assumption is made that functionally equivalent regions from different subjects lie in approximately the same part of the brain. This is usually only partially true, and some smoothing is necessary for statistical analysis. Because there is usually no perfect match for every voxel registered to normalized space, some regions are compressed while others are dilated, which in statistical analysis affects the contribution of a resized region.
Applications and examples
MRI-PET
The epileptogenic zone frequently exhibits decreased glucose metabolism interictally, which often extends beyond the seizure focus (5). Correlation of PET hypometabolism with the underlying cortical anatomy (typically via a high-resolution MRI) is an important step in planning resective epilepsy surgery with concordant FDG-PET hypometabolism, as this metabolic anomaly can significantly influence the location and amount of tissue resected. The FDG-PET coregistred with MRI was shown to be highly sensitive to detect focal cortical dysplasias type II and was also predictive for seizure freedom (37). Figure 9.2 illustrates a patient with right temporal lobe epilepsy confirmed by hypometabolism in the right mesial temporal lobe. The patient’s FDG-PET was registered to a high-resolution seizure protocol MRI via a rigid body mutual information algorithm.
Figure 9.2 F-18 fluorodeoxyglucose PET imaging registered to a high-resolution postoperative T1-weighted MRI for a patient with right temporal lobe epilepsy. The hypometabolic defect identifies the anterior temporal lobe as abnormal despite the absence of a structural lesion in the original preoperative MRI.
MRI-SISCOM and comparison to STATISCOM
Subtraction ictal SPECT coregistered to MRI (SISCOM) (38) uses voxel-based rigid body registration to register a patient’s interictal SPECT image to the ictal SPECT; and following normalization, subtraction, and thresholding, to register the activation map to the patient’s high-resolution MRI. Statistical comparison of the difference image to paired sequential SPECT images from neurologically normal volunteers is possible once the normal SPECT images have been nonlinearly registered to the patient’s brain. This process has been shown to improve the sensitivity and specificity of ictal SPECT studies (2,3). Figure 9.3 illustrates Tc-99m ethyl cysteine dimer SISCOM images (top) and the same ictal and interictal SPECT images analyzed statistically against a group of 30 paired SPECT images of normal volunteers for a patient with left temporal lobe epilepsy. Statistical mapping was performed in MATLAB (Mathworks Inc., Natick, MA) using custom software and SPM (Wellcome Trust Centre for Neuroimaging).
Figure 9.3 Subtraction ictal SPECT (top) and statistical mapping (bottom) of the SPECT activations in a patient with left temporal lobe epilepsy.
Electrode coregistration and seizure mapping
Intracranial EEG recordings of seizure activity represent the current “gold standard” of epileptogenic localization for resective surgery. It is often useful to visualize electrode contact positions in conjunction with the underlying cortical anatomy, and it is possible to create visualizations of neuronal activity interpolated on to the cortical surface, providing a spatial context for electrophysiological activity. Electrode positions are typically identified in X-ray CT images taken immediately following electrode implantation. In Figure 9.4, registration of the CT (B) into the space of the patient’s high-resolution MRI (A) was accomplished using a normalized mutual information algorithm implemented in Analyze (Biomedical Imaging Resource, Mayo Foundation, Rochester, MN). Averaging the two images is shown in (C). The cortical surface in the MRI was identified using a semi-automated threshold-morphology algorithm, and three-dimensional renderings were created using a volume compositing algorithm. Electrode positions are identified by the red spheres, with the resected cortical tissue, identified by registering and subtracting the patient’s postresection MRI, shown in yellow. Time-frequency analysis of one of the patient’s recorded habitual seizures was performed, and the electrophysiological activity at 40 Hz midway through the seizure was color-coded and mapped on to the cortical surface. This electrophysiological mapping can be an important step for accurate clinical diagnosis and treatment as well as for scientific electrophysiological research.
Figure 9.4 Registration of a patient’s high-resolution structural MRI (A) with electrode implant CT (B) allows the fused cortical surface and electrodes (C) to be rendered in three dimensions, illustrating the relative locations of electrode contacts with the cortical anatomy and eventual resection site, shown in yellow (D). The electrode registration also permits mapping of recorded electrophysiology (activity at 40 Hz during a seizure) on to the cortical surface (E).
Neuroimaging tools
Most of the previously mentioned software packages are freely available via the internet and are able to run on various operating systems (36) depending on preference of the user. A moderately powerful workstation is recommended, although for single-case analysis the difference in computing time between a high-end and mid-tier workstation is usually negligible. In a large data set, total computing time starts to play a more important role and use of distributed computing and/or general-purpose computing on graphical processing units (GPGPU) significantly shorten the analysis. For further imaging examples and sources the reader is directed to the website of Neuroimaging Informatics Tools and Resources Clearinghouse (NITRC, www.nitrc.org) funded by the National Institutes of Health Blueprint for Neuroscience Research.
Conclusion
Image registration is an important tool for correlating functional measurements with their underlying anatomical substrate in preoperative planning for resective epilepsy surgery. Successful surgical management of focal epilepsy depends on accurate identification and delineation of the seizure onset zone for resection. In the absence of an MRI-visible structural lesion, epileptogenic localization depends on functional information from imaging modalities including PET, SPECT, and functional MRI, as well as scalp and intracranial electroencephalographic recordings of epileptiform discharges. In developing the clinical localization hypothesis that will guide surgical resection, information from these different modalities’ images must be compared and analyzed with respect to one another and the patient’s underlying cerebral anatomy, and this integration mandates coregistration of these functional modalities into a common physical coordinate system. Dysfunctional patterns not otherwise apparent may be highlighted by statistical comparison of a patient’s functional scans with a population of normal subjects, if the patient’s scans and control scans are nonlinearly transformed into a common anatomical reference space.
References
Findit@CUHK Library | Google Scholar | PubMed
Findit@CUHK Library | Google ScholarRelated posts:
Chapter 1 – Scope and implications of MRI-negative refractory focal epilepsy
Chapter 2 – Seizure semiology and scalp EEG in MRI-negative refractory focal epilepsy
Chapter 21 – Conclusion
Chapter 7 – Electric source imaging in MRI-negative refractory focal epilepsy
Chapter 20 – Neuropsychological issues in MRI-negative focal epilepsy surgery: evaluation and outcomes
Chapter 19 – Histopathology findings in MRI-negative focal epilepsy
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
