The complexity of human brain structure mandates the use of engineering approaches drawn from computer vision, image analysis, computer graphics, and artificial intelligence research fields to manipulate, analyze, and communicate brain data. The rapid growth in brain imaging technologies has also been matched by an extraordinary increase in the number of investigations analyzing brain structure and function in clinical and research settings.
Image registration is central to many of the challenges in brain imaging today. Initially developed as an image processing subspecialty to geometrically transform one image to match another, registration now has a vast range of applications. In this chapter, we review the registration strategies currently used in medical imaging, with a particular focus on their ability to detect and measure differences. These include methods developed for automated image labeling and for pathology detection in individuals or groups. We show that these algorithms can serve as powerful tools to investigate how regional anatomy is altered in disease, and with age, gender, handedness, and other clinical or genetic factors. Registration algorithms can encode patterns of anatomic variability in large human populations and can use this information to create disease-specific, population-based brain atlases. They may also fuse information from multiple imaging devices to correlate different measures of brain structure and function. Finally, registration algorithms can even measure dynamic patterns of structural change during brain development, tumor growth, or degenerative disease processes [107,116,120].
Copyright © 2000 by Academic Press.
All rights of reproduction in any form reserved.
Normal anatomic complexity makes it difficult to design automated strategies that detect abnormal brain structure. Considerable research has focused on uncovering specific patterns of anatomic alterations in Alzheimer's disease (AD) or other dementias , schizophrenia [25,68], epilepsy , attention deficit hyperactivity disorder (ADHD; ), autism [24,40], and cortical dysplasias . At the same time, brain structure is so variable that group-specific patterns of anatomy and function are often obscured. Reports of structural differences in the brain linked to gender, IQ, and handedness are a topic of intense controversy, and it is even less clear how these factors affect disease-specific abnormalities . The importance of these linkages has propelled registration to the forefront of brain imaging investigations. To distinguish abnormalities from normal variants, a realistically complex mathematical framework is required to encode information on anatomic variability in homogeneous populations . As we shall see, elastic registration or warping algorithms offer substantial advantages for encoding patterns of anatomic variation and detecting pathology.
One of the driving forces that made registration important in brain imaging was the need to perform brain to brain comparisons. Anatomic variations severely hamper the integration and comparison of data across subjects and groups [81,129]. Motivated by the need to standardize data across subjects, registration methods were developed to remove size and shape differences that distinguish one brain from another . Transforming individual datasets into the shape of a single reference anatomy, or onto a 3D digital brain atlas, removes subject-specific shape variations and allows subsequent comparison of brain function between individuals [4,15,64,91]. For similar reasons, deformable brain atlases are based on the idea that a digital brain atlas can be elastically deformed to fit a new subject's anatomy [15, 39, 49, 62,93,95,121]. High-dimensional brain image registration or warping algorithms [13,15,16,19,51,59,90,106,111,125] are discussed in detail in this chapter. These algorithms can transfer 3D maps of functional and vascular territories onto the scan of any subject, as well as information on tissue types, cytoarchitecture, and histologic and neurochemical content .
As a valuable by-product, 3D warping algorithms also quantify local and global shape changes. The complex profiles of dilation and contraction required to warp an atlas onto the new subject's brain provide an index of the anatomical shape differences between that subject's brain and the atlas [5,27,114]. Differences in regional shape can be assessed by the displacement required to locally deform one brain volume into another, and can be further examined by applying vector and tensor field operators to the transformation field [108,115]. As a result, deformable atlases not only adapt to individual anatomy, but offer a powerful strategy to analyze developmental, age-related, or pathologic variations.
Pathology detection algorithms are discussed later (Section 4). These invoke deformation fields that match one brain with a large number of others. The result is a probabilistic brain atlas that encodes patterns of anatomic variation in human populations, and incorporates algorithms to detect structural variants outside of the normal range [78,114].
When applied to scans acquired over many months or years from a single subject, 3D warping algorithms can also calculate measures of local and global shape change over time [115,116,120]. In many ways, static representations of brain structure are ill suited to investigating dynamic processes of disease. With warping algorithms, measures of dilation rates, contraction rates, and rates of shearing and divergence of the cellular architecture may be computed locally, for all structures, directly from the deformation field that matches one scan with the other. As a result, warping algorithms offer a powerful strategy to track temporal change and classify age-related, developmental, or pathologic alterations in anatomy.
The challenges created by cross-subject variations in brain structure prompted us to explore hybrid approaches for brain image registration and pathology detection. In these approaches, computer vision algorithms and statistical pattern recognition measures are integrated with anatomically driven elastic transformations that encode complex shape differences in anatomy. As a result, objective criteria can be defined to identify how brain structure is altered by age, gender, handedness, disease, and other genetic or demographic factors . We begin with an overview of 2D and 3D image warping algorithms widely used in brain imaging, in the hope that hybrid algorithms will be developed in the future that capitalize on the merits of each approach.
Was this article helpful?