Analysis of automated methods for spatial normalization of lesioned brains

Abstract

Normalization of brain images is a crucial step in MRI data analysis, especially when dealing with abnormal brains. Although cost function masking (CFM) appears to successfully solve this problem and seems to be necessary for patients with chronic stroke lesions, this procedure is very time consuming. The present study sought to find viable, fully automated alternatives to cost function masking, such as Automatic Lesion Identification (ALI) and Diffeomorphic Anatomical Registration using Exponentiated Lie algebra (DARTEL). It also sought to quantitatively assess, for the first time, Symmetrical Normalization (SyN) with constrained cost function masking. The second aim of this study was to investigate the normalization process in a group of drug-resistant epileptic patients with large resected regions (temporal lobe and amygdala) and in a group of stroke patients. A dataset of 500 artificially generated lesions was created using ten patients with brain-resected regions (temporal lobectomy), ten stroke patients and twenty five-healthy subjects. The results indicated that although a fully automated method such as DARTEL using New Segment with an extra prior (the mean of the white matter and cerebro-spinal fluid) obtained the most accurate normalization in both patient groups, it produced a shrinkage in lesion volume when compared to Unified Segmentation with CFM. Taken together, these findings suggest that further research is needed in order to improve automatic normalization processes in brains with large lesions and to completely abandon manual, time consuming normalization methods.

Introduction

Spatial normalization is one of the most important steps in second-level group magnetic resonance imaging (MRI) analyses. Structural images of participants are normalized to a template (standard or group), ensuring that a one-to-one correspondence among the brains of each individual in the group is created. Normalization becomes more complex when it has to deal with patients with brain lesions. These brains have often greater differences than those individual variations characterizing healthy brains due to important lesions or pathologies (Brett et al., 2001). Correct normalization of individual brains is essential to ensure that brain areas are properly aligned, maximizing sensitivity and minimizing false-negative results. To this end, multiple normalization algorithms have been implemented in fully automated software programs.

Two of the most used normalization algorithms are the Diffeomorphic Anatomical Registration using Lie Algebra (DARTEL) (Ashburner, 2007) and its predecessor, Unified Segmentation (Ashburner and Friston, 2005), implemented in the Statistical Parametric Mapping software (SPM, Wellcome Department of Imaging Neuroscience, University College, London, UK, www.fil.ion.ucl.ac.uk/spm/). Unified Segmentation combines segmentation, bias correction and spatial normalization under the same iterative model using white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) tissue maps as priors (TPMs). These TPMs are deformed by a linear combination of a thousand cosine transform bases, and several Gaussian distributions are used to model the intensity of each tissue class. Unlike Unified Segmentation, DARTEL utilizes a large deformation framework to preserve topology, assuring that the deformations are invertible, diffeomorphic and parameterised by a flow field. Rather than using a thousand parameters for the registration process as Unified Segmentation, DARTEL uses about six million and the registration itself involves alternating between computing an average template of the GM and WM TPMs from all subjects and warping all subjects' TPMs into a better alignment with the template created (Ashburner, 2009). Both of these algorithms are segmentation-dependant, as DARTEL needs the segmentations of all subjects in the group to create the average template and Unified Segmentation combines segmentation with normalization. Besides Unified Segmentation, another way to provide DARTEL with the GM and WM segmentations it needs is the New Segment toolbox under the SPM8 distribution. This algorithm is essentially the same as that described in the Unified Segmentation model, except for a different treatment of the mixing proportions, the use of an improved registration model, the ability to use multi-spectral data and an extended set of tissue probability maps (The FIL Methods Group, 2010). The default set includes TPMs for gray matter, white matter, CSF, bone, soft tissue and air/background, but allows the user to define as many extra TPMs as desired (The FIL Methods Group, 2010). New Segment can also provide deformation fields which can be later used to spatially normalize images, but in this manuscript it has been only used to provide DARTEL with the segmentations it needs.

In addition, as part of the Advanced Normalization Tools (ANTS) (Avants et al., 2011a), the Symmetric normalization algorithm (SyN) (Avants et al., 2008), also based on large deformations, has shown to perform at least as good as DARTEL when dealing with healthy subjects (Klein et al., 2009). SyN keeps symmetry when connecting two images in a geodesic (shortest distance in space) link, meaning that the path from A to B is the same than the one from B to A, irrespective of the optimisation or the similarity metric used (Avants et al., 2008). SyN has about 28 million degrees of freedom (Klein et al., 2009) and uses a gradient optimization scheme which is basically an iteration over time of three steps: computing the similarity gradient, updating the deformation field and regularizing the deformation field. Within ANTS, SyN can work with different similarity metrics as cross-correlation, mutual information or mean square difference (Avants et al., 2011b) and can use different types of regularization based on Gaussian or Bsplines. ANTS also provides optimal template construction and image segmentation, among other things.

In healthy subjects these normalization methods function optimally but, in contrast, spatial normalization suffers from some limitations when normalizing images of patients with large lesions, such as those found in stroke patients (Andersen et al., 2010) or in patients with tumors, cortical dysplasia or atrophy (Crinion et al., 2007). Different procedures have been used when trying to normalize abnormal brains. Initially, cost function masking (CFM) with standard SPM normalization was proposed for routine use when normalizing brains with regions containing abnormal signal intensities (Brett et al., 2001). Most normalization methods calculate a cost function, a measure of the signal intensity difference between a source image and a template, which has to be minimized (Brett et al., 2001). CFM is based on creating a binary mask of the lesioned area and taking the signal under the masked area out of the calculation of the transformations needed to normalize the image. Later, Crinion et al. (2007) proposed that, although for low regularization the use of the Unified Segmentation model with CFM provided a better registration than Unified Segmentation without CFM, when using medium regularization the use of CFM did not improve normalization. These results were assessed in a set of ten patients with different types of brain injuries (including tumor, stroke, cortical atrophy and dysplasia) and the regularization process was formulated as the precision of the bending energy priors on the deformation relative to the squared difference between the observed and normalized images under the Gaussian definition of noise. Nevertheless, a very recent study by Andersen et al. (2010) using a database of 49 chronic stroke patients showed that Unified Segmentation and medium regularization with CFM yielded better normalization results than those of Unified Segmentation with medium regularization alone, demonstrating the need for CFM when dealing with large lesions.

Interestingly, a different approach was recently presented by Seghier et al. (2008) in which adding an extra fourth tissue prior, which was defined as the mean of the cerebrospinal fluid (CSF) and white matter (WM) tissue probability maps (provided by SPM), improved the segmentation using the unified model. Unified Segmentation is computed twice: first, an iteration is computed with the predefined extra class, and then a second iteration is calculated with an updated definition (subject-specific) of the extra class (Seghier et al., 2008). Following this work, the Automated Lesion Identification (ALI) toolbox was developed, which allows the user to implement this type of normalization while also being capable of automatically identifying and delineating lesions.

Although spatial normalization of abnormal brains has been assessed with small-deformation methods such as Unified Segmentation, it has not yet been studied with large deformation techniques such as DARTEL, which have proven to achieve a better normalization in healthy volunteers (Klein et al., 2009, Tahmasebi et al., 2009, Yassa and Stark, 2009). Alternatively, the SyN method allows the possibility to use a technique called constrained cost function masking (CCFM) in order to normalize lesioned brains (Kim et al. 2007), which has been reported to give optimal results when normalizing a group of stroke patients (Schwartz et al., 2009). As in CFM, a lesion mask must be depicted and applied to the registration. In CCFM, the velocity field parameters calculated by SyN under the mask are treated as unknown values and estimated using the information given by the velocity fields of the healthy tissue near the masked area.

In addition to the methodological issues previously discussed, there is a limitation in the generalization of the results obtained by previous studies to all brain lesions. It is important to note that most of the previous analysis dealt with patients with a variety of brain injuries, especially with lesions derived from vascular events, but none of them directly investigated these issues with a set of patients with large brain resections. After surgery, CSF mostly invades the empty space left by the resected tissue; however, the remaining brain tissue also tries to occupy this new empty volume. These variations produce a great variety of morphological brain changes, which might create serious difficulties when trying to normalize images of these injured brains. Because many studies involve tumor-resected patients or epileptic patients with removal of the epileptic focus, these problems are not infrequent (Cheung et al., 2009, Immonen et al., 2010, Yogarajah et al., 2010). Achieving an optimal normalization for these patients is crucial, especially for studies aiming to compare pre- and post-surgery brains.

The first aim of the present study was to evaluate which of the previously described approaches is more effective for abnormal brain normalization. We are particularly interested in automated methods that could lead to the abandonment of manual lesion depicting methods such as CFM or CCFM, which are extremely time consuming as the lesions must be defined manually by an expert neurologist or neuroradiologist and might also be subject to expert biases. We expected that the more sophisticated approach given by DARTEL (Klein et al., 2009, Tahmasebi et al., 2009, Yassa and Stark, 2009) and SyN (Klein et al., 2009), which has proven to provide a better normalization than other traditional SPM based methods in healthy subjects, would achieve a more accurate normalization. We also expected that using an extra class as a prior would cause the ALI toolbox to improve normalization results compared to regular Unified Segmentation. The second aim of this study was to determine the impact of the type of lesion in the performance of the normalization methods, especially when dealing with brain-resected patients. With this purpose in mind, normalization methods were tested on three different groups of participants (healthy patients, stroke patients and epileptic patients with brain-resected regions) to study whether there are differences in normalization performance of the different methods when dealing with different types of lesions.