Guest Editorial Special Issue on Computational ... - IEEE Xplore

2 downloads 114239 Views 561KB Size Report
Oxford, U.K., in 1993, the M.Sc. degree in computer science and the Ph.D. degree ... Women's Hospital and Harvard Medical School, and Associate Professor of ...
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 11, NOVEMBER 2007

1425

Guest Editorial Special Issue on Computational Diffusion MRI

T

HIS special issue of the IEEE TRANSACTIONS ON MEDICAL IMAGING is devoted to the topic of computational diffusion magnetic resonance imaging (MRI). Diffusion MRI has exploded in the last decade since the emergence of diffusion tensor MRI in the mid 1990s. The technique provides a unique probe into the microstructure of living tissue and other materials. In fibrous tissue, diffusion MRI provides estimates of fibre orientations in each image voxel and thus allows tractography algorithms to reconstruct global fibre trajectories and infer connectivity. In vivo connectivity mapping using diffusion MRI provides fundamental neuroscientific insights into the anatomy of the brain. The variety of clinical applications is expanding rapidly and includes detection of lesions and damaged tissue, monitoring of development and disease progression, prognosis of functional impairment, and neurosurgical planning. Early diffusion MRI research focussed mainly on the physics of diffusion in tissue and how it affects the MR signal. However, as the field has become more popular and mature, it has relied increasingly on sophisticated computational techniques. Computational techniques are now essential for addressing issues at each stage of the diffusion MRI pipeline: acquisition, reconstruction, modelling, and model fitting, image processing, fibre tracking, connectivity mapping, visualization, group studies, and inference. The unique images the technique provides present new and compelling challenges to the medical image analysis community. This special issue provides a taste of those challenges and a snapshot of the approaches that the community currently adopts. The work contained in this special issue touches on all stages of the pipeline above. Although several standard diffusion MRI acquisition protocols are now available and widely used, acquisition design is still an issue and careful modelling and optimization can dramatically improve results. A timely concern is how best to exploit multiple coils in parallel imaging. Beaudoin et al. construct a method to improve the combination of signals from multiple array coils and evaluate the stability of standard diffusion MRI markers using their approach. The precise mechanisms of water displacement and relaxation that give rise to the diffusion MRI signal remain poorly understood. Several papers in this issue perform fundamental investigations into models of those mechanisms and what they can estimate. Lätt et al. investigate the ability of the full-width halfmaximum of the displacement distribution, assuming the short gradient pulse approximation, to predict the size of restricted compartments in live brain tissue. Simulations and phantom exis periments suggest that, for achievable gradient pulses, 10 a lower limit on the compartment size we can measure. Peled models white matter as a three compartment system in which

water diffuses, relaxes, and exchanges between an intracellular compartment, its myelin sheath, and an extracellular compartment. Diffusion simulations within this synthetic environment provide important insights into the measured signal and explain an earlier observation that adding an isotropic baseline tensor to the standard diffusion tensor model significantly improves the fit to measured data. Miller et al. also use a diffusion simulation to validate diffusion MRI measurements of hyperpolarized gas in the rabbit lung. They construct environments for the simulation from histological images of the lungs of rabbits with emphysema. Reconstruction of the particle displacement density, and subsequently the fiber orientation distribution, from diffusion-weighted measurements remains an active area of research in diffusion MRI. Diffusion tensor MRI has the well documented limitation that it cannot resolve multiple fiber populations within single voxels. That observation lead to the development of a variety of multiple-fiber reconstruction algorithms over the last five years. Jian and Vemuri examine various implementations of one class of multiple fiber reconstruction: spherical deconvolution algorithms. They find a nonnegative least-squares approach most effective in experiments on both synthetic and rat-brain data. Multiple-fiber reconstructions generally require high quality data. In clinical applications, time is often limited and data quality may be low. Fillard et al. propose techniques for fitting the simple diffusion tensor model to data with low signal to noise. Their fitting routines use the Rician noise model and they incorporate spatial smoothing using a log-Euclidean metric to improve results. Applications of medical imaging techniques often rely on statistical analysis. An important question in diffusion MRI is how to perform statistical analysis to allow comparison and evaluate differences between single voxels or groups of measurements. Kindlmann et al. present a general framework for analyzing diffusion tensors, which separates shape and orientation parameters and defines natural invariant systems. They define third and fourth-order tensors that express the gradients and covariances of diffusion tensors, respectively, and derive relationships with invariant gradients and covariances. Peyrat et al. propose a computational framework for building statistical atlases of fiber architecture from diffusion-tensor images. They exploit recent statistical techniques based on the Riemannian geometry of the space of symmetric positive definite diffusion tensors. They use the framework to construct an atlas of the canine heart. Fonteijn et al. address a different statistical problem. Probabilistic tractography requires estimates of the uncertainty in fiber orientation estimates in each voxel. Markov chain Monte Carlo (MCMC) methods are useful for this purpose and Fonteijn et al. show how to use MCMC in conjunction with multiplefiber reconstruction algorithms (they use the Q-ball algorithm)

Digital Object Identifier 10.1109/TMI.2007.909944 0278-0062/$20.00 © 2007 IEEE

1426

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 11, NOVEMBER 2007

to estimate the posterior distributions of the fiber orientation estimates they provide. Image segmentation is a natural application of diffusion MRI. Many techniques that use diffusion MRI for segmentation exploit tractography algorithms to segment specific fiber tracts, which can be difficult to outline precisely by hand. Awate and Gee point out that diffusion tensors within fiber tracts lie on low dimensional manifolds in the Riemannian space of positive definite matrices. They exploit this observation for tract segmentation using a fuzzy clustering algorithm. Barmpoutis et al. adopt a similar philosophy. They extend B-spline interpolation and approximation to tensor fields and use the technique for diffusion tensor image segmentation by dividing the image into regions with separate smooth spline approximations. Jonasson et al. propose a five-dimensional (three spatial and two directional) representation of fiber orientation information from diffusion MRI as a basis for image processing operations. The directional information comes from diffusion spectrum imaging and they use their framework to define regularization and segmentation algorithms that can separate crossing fiber pathways. Clayden et al. show how to improve the consistency of tractography-based segmentation by developing and exploiting a model of fiber trajectories, which they use as a prior for selecting and grouping streamlines during segmentation. In a similar way, O’Donnell and Westin use clustering techniques to group similar fiber trajectories and create a high-dimensional white-matter atlas, which helps provide consistent segmentation and identification of specific fiber pathways in individual images. Another way to use diffusion MRI for segmentation is simply to group image voxels with similar diffusion properties, which implies similar microstructure. Freidlin et al. use various model selection techniques to choose the simplest from a hierarchy of diffusion models that captures the variability in the data. In this way, they divide voxels into classes with different model complexity, which reflects the complexity of the tissue microstructure. Diffusion MRI presents new opportunities and challenges for image registration and spatial normalization. Fiber orientation estimates provide new information to drive matching that is complementary to anatomical image intensities. However, the directional information complicates image warping. Zhang

et al. show that high dimensional spatial normalization increases the significance of differences between patients and normals, in a voxel-based morphometry study of amyotrophic lateral sclerosis, over affine spatial normalization. They show further that matching the full diffusion tensor, with explicit tensor orientation matching during optimization in their high dimensional registration, increases significance over matching scalar anisotropy. Van Hecke et al. use fluid registration with multivariate mutual information to match diffusion tensor images. They obtain good alignment but observe that the high dimensional warp can affect tensor reorientation adversely. In summary, the papers in this special issue highlight the broad range of computational challenges that the promise and popularity of diffusion MRI has exposed. Significant challenges remain in the area and we hope this collection of work will inspire further development and new ideas from the broader medical image analysis community in this exciting and rapidly expanding area. We would like to thank all the researchers that submitted work to this special issue. Many excellent papers do not appear, but will appear in later issues of this or other journals. We would also like to thank all the reviewers, who ensured the high quality of the final set of papers, the IEEE editorial staff, in particular Marjan Marinissen and Jacqueline Wermers, for their invaluable assistance and Max Viergever for his help and support and for giving us the opportunity to create this special issue. Daniel C. Alexander, Guest Editor Department Computer Science University College London London, U.K. WC1E 6BT Tianzi Jiang, Guest Editor Institute of Automation The Chinese Academy of Sciences Beijing, China 100080 Carl-Fredrik Westin, Guest Editor Department of Radiology Harvard Medical School, Brigham and Women’s Hospital Boston, MA USA 02215

Daniel C. Alexander (M’00) received the B.A. degree in mathematics from Oxford University, Oxford, U.K., in 1993, the M.Sc. degree in computer science and the Ph.D. degree in computer vision from University College London, in 1994 and 1997, respectively. He worked as a Postdoctoral Research Fellow at the GRASP Laboratory at the University of Pennsylvania, Philadelphia, until 2000 when he returned to University College London (UCL), London, U.K. for a permanent academic position. He is Reader in Imaging Sciences in the Center for Medical Image Computing and the Department of Computer Science at UCL. His main research areas are computer vision and medical imaging. He has worked with diffusion MRI for nearly ten years making key contributions in tensor warping and registration, multiple-fibre reconstruction, probabilistic tractography, acquisition optimization and model fitting and estimation.

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 11, NOVEMBER 2007

1427

Tianzi Jiang (M’98–SM’04) received the Ph.D. degree in computational mathematics from Zhejiang University, Zhejiang, China, in 1994. After he graduated, he worked as a Postdoctoral Research Fellow (1994–1996) and an Associate Professor (1996–1999), and Full Professor (1999–present) at his current institution. During that time, he worked as a Vice-Chancellor’s Postdoctoral Fellow at the University of New South Wales, a Visiting Scientist at the Max Planck Institute for Human Cognitive and Brain Sciences, a Research Fellow at the Queen’s University of Belfast, and a Visiting Professor at University of Houston. He is a Chief Professor of National Laboratory of Pattern Recognition, Institute Automation, the Chinese Academy of Sciences. He is the Chinese Director of the Sino-French Laboratory in Computer Science, Automation and Applied Mathematics (LIAMA) since 2006. His research interests include anatomical and functional brain imaging, complex brain networks, imaging genetics, and their clinical applications and drug development.

Carl-Fredrik Westin (M’03) received the M.Sc. degree in applied physics and electrical engineering, the Lic. Techn. degree on the topic of feature extraction from a tensor image descriptions, and the Ph.D. degree in computer vision from Linkoping University, Linkoping, Sweden, in 1988, 1991, and 1994, respectively. He joined Brigham and Women’s Hospital and Harvard Medical School, Boston, MA, in 1996. He is Director of Laboratory of Mathematics in Imaging, Department of Radiology, Brigham and Women’s Hospital and Harvard Medical School, and Associate Professor of Radiology at Harvard Medical School. His research interests are focused on medical applications of image analysis. He is currently working on analysis of diffusion MRI data, and automated segmentation and registration of data from MRI, CT, and ultrasound, using multidimensional signal processing techniques.