There are many important questions we want to answer through DTI analysis of AD. For example, what are the AD-specific features that can be observed with DTI? Can these features be observed in persons with mild cognitive impairment, thought to be the earliest symptomatic stage of AD, or even in the presymptomatic phase? How does DTI reflect or predict the progression of AD? Are there correlations between DTI and cognitive functions? To answer these questions, the quantification of DTI parameters is the first important step.
Several issues arise when we quantify DTI results. First, regions of interest (ROI) should be defined to measure DTI-derived parameters, such as FA or MD. The smallest ROI we can define is a single voxel, and the maximum ROI is the whole brain. The localization information is maximized when the smallest ROI (single voxel) is adopted, but the statistical power is lost because of the low signal-to-noise ratio and the difficulty of identifying the corresponding voxel across subjects. The statistical power is maximized when the size and shape of the ROI exactly follow pathological locations. If we have an a priori
hypothesis about the locations of the pathologic tissues, we could pre-define the size and shape of ROIs according to the hypothesis. If we hypothesize that the pathology is seen in specific fiber tracts, we can use tractography to draw ROIs (tract-specific analysis, see, eg. [54
]). ROI-based DTI analyses have been widely used for AD studies and have successfully identified reduced FA or increased MD, or both, within the splenium of the corpus callosum [47
], the cingulum bundle [11
], and the fornix [43
]. However, this approach is hypothesis-dependent, and the majority of the brain area remains unexamined, which makes it difficult to evaluate the localization specificity.
Second, we need to decide whether we should explore the whole brain or limited areas of the brain. Whole brain analysis is ideal for evaluating the regional specificity of the findings. However, drawing a number of ROIs manually that would cover the whole brain is a tremendous effort. Thus, automated method, such as those based on image normalization (transformation), are typically used. After transforming images to a common template space (atlas), we can even quantify the image at the voxel level (voxel-based analysis). Although image normalization has been widely used to analyze conventional MRI contrasts, the transformation of DTI poses a unique challenge. DTI data consists of tensor fields (as opposed to scalar fields for conventional MRI), and white matter tracts revealed by the tensor field must be registered after normalization [63
].To avoid false-positive and false-negative findings, the accuracy of the registration is a crucial requirement. Accuracy is especially critical when dealing with small structures, string-like structures, and sheet-like structures, which are often found in the white matter. In these structures, only a few pixel gaps between the subject and template image will cause significant missregistration.
Various non-linear transformation methods have been proposed for DTI analyses, such as tensor-to-tensor matching [65
], scalar measures matching [69
], or other DTI-derived information, as well as some combination of these methods [71
]. Transformation methods based on non-DTI contrasts have also been applied to transform DTI, with high registration accuracy [64
To perform whole brain analysis with no a priori
hypothesis, voxel-based analysis is one of the most widely used approaches [29
]. is an example of such an analysis, designed to find brain areas with AD-specific white matter alterations. This analysis indicates significant FA reduction in the fornix, the splenium of the corpus callosum, as well as several small areas in the superficial white matter in the frontal lobe. Although this result seems to be consistent with previous ROI-based investigations, we must interpret the results with great caution, since this approach tends to miss the widely distributed regions that show only small changes in the parameters [80
]. One of the attempts to overcome such limitations is to apply multivariate models [81
], such principal component analysis (PCA) [78
] or canonical correlation analysis (CCA) [82
], which have already been applied to DTI analysis of AD. Other methods include voxel grouping. For example, if we hypothesize that the white matter pathology of AD is tract-specific or structure-specific, especially in the early stages, we may apply tract-based voxel grouping, such as Tract-Based Spatial Statistics (TBSS) [83
], or structure-based voxel grouping, such as atlas-based analysis (ABA) [84
]. Indeed, TBSS has already revealed important findings, such as deteriorations in the limbic fibers, the fronto-occipital fasciculi, the inferior longitudinal fasciculi, and the forceps major, even in the early-symptomatic patients or in participants at high risk for developing AD [45
]. ABA is a method that uses a set of pre-defined ROIs, called a parcellation map (), which covers the entire brain, in the atlas space. The parcellation map can be overlaid on the images normalized to the atlas space to measure DTI-derived parameters (eg., FA or MD)in each ROI (parcel), or can be transformed to each image to measure DTI-derived parameters as well as the volumes of each ROI. One additional feature that sets the ABA apart from TBSS is that ABA provides morphometric (volume) information about brain atrophy in the volume of each parcel. Our initial results from ABA () indicate a higher sensitivity for ABA in detecting changes in FA, especially in the areas with widely distributed small FA reductions, compared to the voxel-based analysis. The drawback of this approach is that if the region is limited in the small portion of the structure (parcel), the effect is diluted and sensitivity is decreased.
Fig. 2 Statistical group comparison of FA after image normalization. Nineteen patients with Alzheimer’s disease and 22 age-matched cognitively normal participants were compared, and the areas with significance (t-test, p < 0.05 after correction (more ...)
Fig. 3 Atlas-based analysis. The original FA map from a patient with Alzheimer’s disease (A) was normalized to the atlas space (B). The atlas used as the template is shown in (C). After image normalization, pre-defined three-dimensional ROIs (color-contours, (more ...)