Multimodal Deformable Registration of Traumatic Brain Injury MR Volumes via the Bhattacharyya Distance

Description

An estimated 1.7 million Americans sustain traumatic brain injuries (TBI's) every year. The large number of recent TBI cases in soldiers returning from military conflicts has highlighted the critical need for improvement of TBI care and treatment, and has drawn sustained attention to the need for improved methodologies of TBI neuroimaging data analysis. Neuroimaging of TBI is vital for surgical planning by providing important information for anatomic localization and surgical navigation, as well as for monitoring patient case evolution over time. Approximately 2 days after the acute injury, magnetic resonance imaging (MRI) becomes preferable to computed tomography (CT) for the purpose of lesion characterization, and the use of various MR sequences tailored to capture distinct aspects of TBI pathology provides clinicians with essential complementary information for the assessment of TBI-related anatomical insults and pathophysiology.

Image registration plays an essential role in a wide variety of TBI data analysis workflows. It aims to find a transformation between two image sets such that the transformed image becomes similar to the target image according to some chosen metric or criterion. Typically, a similarity measure is first established to quantify how "close" two image volumes are to each other. Next, the transformation that maximizes this similarity is typically computed through an optimization process which constrains the transformation to a predetermined class, such as rigid, affine or deformable. Numerous challenges associated with the task of TBI volume co-registration can exist if data acquisition is performed multimodally, and additional complexities can also arise due to the large degree of algorithmic robustness that may be required in order to properly address pathology-related deformations. Because the deformation of patient anatomy and soft tissues cannot typically be represented by rigid transforms, image registration often requires deformable image registration (DIR), i.e., the necessity of applying nonparametric infinite-dimensional transformations.

We propose to replace the Mutual Information (MI) criterion for registration with the Bhattachayya distance [1] for multimodal DIR problem. The advantage of BD over MI is the superior behavior of the square root function compared to that of the logarithm at zero, which yields a more stable algorithm. We use a viscous fluid model [2], which takes into account the physical models of tissue motion to regularize the deformation fields and also involves free-form deformation. On the other hand, the DIR algorithm is computationally expensive when implemented on conventional central processing units, which can be detrimental particularly when three-dimensional (3D) volumes-rather than 2D images-need to be co-registered. In clinical settings that involve acute TBI care, the amount of time required by the processing of neuroimaging data sets from patients in critical condition should be minimized. To meet this clinical requirement, we have implemented our algorithm on a graphics processing unit (GPU) platform [3].

Results

Before applying our deformable registration algorithm, all image volumes were co-registered by rigid-body transformation to the pre-contrast T1-weighted volume acquired during the acute baseline scanning session. All image volumes were co-registered by rigid-body transformation to the pre-contrast T1-weighted volume acquired during the acute baseline scanning session. This approach is helpful when accounting for head tilt and when attempting to reduce errors in computing the local deformation fields. Another operation which is applied before performing the registration is skull stripping. Skull stripping is an important preprocessing step in many neuroimaging analyses. This is particularly useful in our case because images acquired for some modalities exhibited appreciably more extracranial swelling compared to others. Without skull stripping, the registration algorithm would try to match the outer boundary. Also having a brain mask will shorten the computational time. It should be noted that it is not necessary to have a very accurate brain mask because the brain is the main structure of interest in TBI image analysis. Because all modalities are co-registered to T1, one only needs to perform skull stripping once, i.e. on the T1 volume.

Excluding preprocessing steps, the registration of two volumes of sizes 256x256x60 is found to require 6 seconds on the GPU. For comparison, BRAINSFit [4], NiftyReg [5] and ANTs [6] were tested on the same data under identical hardware configurations. BRAINSFit and NiftyReg are both B-spline based implementations, while ANTs implements a diffeomorphic deformation. BRAINSFit is used here as a plugin in 3D Slicer, and is found to require 8 minutes. NiftyReg contains both CPU and GPU implementations and it takes 15minutes and 35 seconds (CPU) and 2 minutes 18 seconds (GPU). ANTs, a diffeomorphic registration method, requires one hour and 4 minutes. The visual results are illustrated below.

It is worth noticing that the topology changes around a hemorrhaging white matter region in MR images of patient #2. The structure is ambiguous in the corresponding area of T1 due to the inability of the latter technique to distinguish between edemic and nonedemic tissue, on one hand, and between hemorrhagic and non-hemorrhagic edema, on the other hand. The MR images as deformed by various registration methods have different visual features. All these methods assume that the deformation field is smooth; when this assumption is not valid, each method imposes its own interpretation of the “correct” deformation. Though important, it is difficult to design registration algorithms which can handle topological changes, especially for registering TBI data across time. This is beyond the scope of this paper and will be our future research direction.

The hyper-intense region in the frontal lobe on the left of the image is an intra-cerebral, CSF-filled cavity whose presence is due to TBI-related loss of white matter tissue between the acute and chronic time points. Although the chronic stage of TBI is typically associated with partial recovery of function in patients such as the one selected here, brain regions subjected to large primary injuries are often found to lose an appreciable amount of tissue longitudinally. Consequently, the example we present illustrates how brain topology can actively change with time after TBI, which highlights the complexities associated with developing registration methods for this condition.

Clinical Significance

Robust methods for TBI assessment can play an essential role during both acute and chronic therapy of this condition for several reasons. Firstly, it is well appreciated that proper management of TBI sequelae can alter their course, improve mortality and morbidity, reduce hospital stay and decrease health care costs. Consequently, TBI neuroimaging is important for surgical planning by providing useful information for guiding decisions concerning the aggressiveness of TBI treatment. Nevertheless, the use of automatic MRI image processing algorithms to the clinical assessment of TBI cases remains problematic due to the fact that many existing methods are insufficiently robust from the standpoint of accurately capturing TBI-related changes in brain anatomy. Thus, although significant recent progress has been achieved in the development of image analysis tools, TBI quantification remains difficult, particularly for the purpose of improving clinical outcome metrics.

The two figures below show the norms of the deformation fields and their 2D motion grid for both the acute and chronic stages. For T2, FLAIR and GRE volumes, the largest amount of deformation is observed bilaterally in the deep periventricular white matter, possibly as a consequence of hemorrhage and/or CSF infiltration into edemic regions, which can alter voxel intensities in GRE and FLAIR imaging, respectively. In the case of DWI, notable deformation is observed frontally and frontolaterally; in the former case, this may be the result of warping artifacts due to the large drop in the physical properties of tissues at the interfaces between brain, bone and air. In the latter case, the deformation is possibly due to the presence of TBI-related edema, which can substantially alter local diffusivity values. Similar effects due to these causes are observed with DWI and with perfusion imaging in both acute and chronic scans.

Our visual representations involving deformation norms and motion grids can be interpreted as error maps, which allow one to acquire improved understanding of how pathology discriminately affects each imaging modality. This may be useful in order to ascertain how various phenomena associated with TBI pathology (e.g. the deformation norm of CSFperfused edema in the case of FLAIR or hemorrhage in the case of GRE imaging) can affect brain shape and tissues. For example, large deformation field norms associated with periventricular pathology at the acute time point (as obviated in Figure 7) can be conceptualized as quantitative indicators regarding the spatial extent and manner in which aspects of the pathology that are captured by each of these techniques affect brain tissues differentially.

Reference

1.Yifei Lou, Andrei Irimia, Patricio Vela, Micah C. Chambers, Jack Van Horn, Paul M. Vespa and Allen Tannenbaum. Multimodal Deformable Registration of Traumatic Brain Injury MR Volumes via the Bhattacharyya Distance. Submitted to IEEE Transactions on Bioengineering, 2012

2. Yifei Lou, Xun Jia, Xuejun Gu and Allen Tannenbaum. A GPU-based Implementation of Multimodal Deformable Image Registration Based on Mutual Information or Bhattacharyya Distance. Insight Journal, 2011.

3. E. D’Agostino, F. Maes, D. Vandermeulen, and P. Suetens. A viscous fluid model for multimodal non-rigid image registration using mutual information,” MICCAI, 2002, pp. 541–548

4. H. Johnson, G. Harris, and K. Williams, “BRAINSFit: Mutual information registrations of whole-brain 3D images, using the insight toolkit,” http://hdl.handle.net/1926/1291, 2007.

5. M. Modat, G. R. Ridgway, Z. A. Taylor, M. Lehmann, J. Barnes, D. J. Hawkes, N. C. Fox, and S. Ourselin, “Fast free-form deformation using graphics processing units,” Computer Methods and Programs in Biomedicine, vol. 98, no. 3, 2010.

6. B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee, “Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain,” Med Image Anal, vol. 12, pp. 26–41, 2008.