Creation of WMPM and usage
Similar to cortical atlases, there is a certain degree of arbitrariness in the definition of the boundaries of the WMPM because many anatomical entities, such as the “corona radiata” and the “corpus callosum,” often do not have clear tissue boundaries. Therefore, the WMPM should be considered a guide for evaluating white matter anatomy rather than a gold standard for anatomical definition. There are several ways to use the WMPM. For example, if one is interested in studying white matter lesions, such as those occurring in multiple sclerosis or stroke patients, we often need to identify, report, and compare the lesion locations with those in other patients and correlate them with functional deficits (structure-function analyses). Template-based stereotaxic coordinates after brain normalization are widely used for these purposes. In this conventional approach, lesion locations are expressed as three-dimensional standardized coordinates, in which each anatomical coordinate is treated as an independent entity. The WMPM can add another anatomical dimension by grouping voxels that belong to specific white matter structures; for example, two lesions in two patients with different normalized coordinates may belong to the same white matter tract. This new anatomical dimension added by the WMPM may increase the sensitivity and specificity of group analyses, such as identification of white matter tracts that are most sensitive to disease or involved in specific functional deficits.
WMPM-based quantification and registration quality
In order to measure MR parameters, such as FA, ADC, T2, or magnetization transfer ratio (MTR), manual ROI definition is one of the most widely adopted approaches. Although this is a valid approach, it has several drawbacks. First, it is usually hypothesis-driven, in which target brain regions and control regions are pre-selected based on expectation. Comprehensive analyses of the entire brain using multiple 3D ROIs may be possible, but would be too time-consuming for practical use. Second, the reproducibility of the manual delineation is often a subject of criticism. The pre-parcellated WMPM provides us with a means to evaluate the large number of white matter structures automatically and reproducibly, which could be a useful tool for initial whole-brain screening to assess the status of the brain and bring our attention to sensitive brain regions for more refined investigation.
In this paper, the WMPM was used for measuring regional intensities (i.e., FA) in two different ways; it was used as a guide for manual ROI drawing (manual and hybrid methods) or automated parcellation (Automated I and II). The high inter-rater reproducibility of the manual and hybrid approaches (κ > 0.85 and CoV < 3%) suggests that it is an effective guide for ROI drawing. With the hybrid approach, inter-rater reproducibility is also high (κ > 0.75 and CoV < 3%), which is attributable to the elimination of variability in slice selections among raters. The advantages[SM23] of the hybrid approach include: 1) it can correct differences in brain orientations and, thus, extracted slices are likely to be more consistent across subjects; 2) objective criteria (i.e., the coordinates) for slice identification makes the ROI drawing process easier; and 3) the slice and ROI locations can be reported using a widely used coordinate system, such as ICBM-152.
In[SM24] this study, we did not include the results of a manual ROI approach without using the WMPM as guidance. Usually, we need to determine some type of pre-defined (often visual) protocols to define ROIs. Without such protocols, the reproducibility of the definition of the border for some white matter tracts becomes very poor; for example, the corpus callosum in an axial or a coronal slice is often a continuous entity and different operators may use different anatomical clues to define the border. The WMPM can be considered one of the pre-defined, 3D, ROI drawing protocols in this regard. Our software, Landmarker and RoiEditor, provides interfaces for the brain normalization and the WMPM-guided ROI drawing.
Using the hybrid method as a reference, the accuracy of the automated methods was evaluated (). Although one of the advantages of the automated methods is three-dimensional WMPM analysis, the comparison was limited to a representative 2D axial slice (z = 79 or y = 104) because it would be too time-consuming to manually define multiple (26 regions in this paper) 3D ROIs. Both the linear (Automated I) and non-linear (Automated II) methods show high correlation (r2 > 0.94) for the FA values of the 26 anatomical regions. The Automated I method, however, has a large standard deviation among the normal population for several white matter tracts. The tracts with the highest variability are the right and left cingulum, which are small tracts, and a slight mis-registration can lead to significant inaccuracy.
While the rapid and three-dimensional quantification by WMPM is a significant advantage over manual-based analyses, the drawback is that the accuracy depends on the quality of image normalization, which is often known to be inaccurate; if registration is poor, the WMPM would not align to the white matter structures of the subject. To measure the quality of structural alignment, we used landmark distances between the template and normalized subject data. The[SM25] ICBM-DTI-81, which is based on normal population averages, provides registration errors mostly less than 3 mm. This registration quality may sound unexpectedly high, compared to previous reports based on cortical registration (Salmond et al., 2002
; Thompson and Toga, 1996
; Van Essen and Drury, 1997
). However, this result is in line with previous registration studies measuring deep brain structures (Ardekani et al., 2005
; Grachev et al., 1999
We would like to emphasize that the registration quality measurements in this study are based on normal adult subjects and do not represent patients with significantly altered neuroanatomy. Compromised neuroanatomy in patients, such as enlarged ventricles, often cannot be normalized by linear transformation. In this case, the registration quality of the WMPM is expected to deteriorate. It is, therefore, very important to carefully interpret the results of automated MR intensity measurements. If abnormalities, such as reduced FA, are found in certain white matter regions, this could be due to anatomical changes and subsequent poorer registration in such areas. Visual inspection of registration quality and reexamination by manual ROI of such abnormal regions are recommended. If poor registration is the reason for the abnormal intensity values (e.g., decreased FA), it implies consistent anatomical differences, but not FA differences, in the abnormal area. In this case, size measurements of the putative structure may be advisable.
In this study, 4th order polynomial transformation was used for non-linear transformation, which improved variability among the normal population observed with the linear normalization. However, the 4th order transformation may not be elastic enough to remove large anatomical differences often observed in patient groups. To ensure better registration qualities, non-linear transformation with higher elasticity will be an important future effort. To fully exploit the advantages of high-order non-linear transformation, however, the population-averaged template may need to be recreated, because the ICBM-152 and ICBM-DTI-81, being obtained by linear normalization, do not have clear anatomical definition as a target of such transformation methods.
One[SM26] important question that remains unanswered in this paper is the effect of age. In this paper, we pooled data from subjects from 18 – 59 years of age, assuming that the white matter anatomy is not significantly different among these age groups. If this assumption does not hold true, the precision of the atlas could be increased by creating multiple atlases at each age range. Similarly, it remains an important question whether our atlas can be applied to subjects older than 60 or younger than 18 years of age. These issues are also related to the accuracy of the normalization procedure; if the aged-dependent differences can be removed by the transformation method of choice, the impact of age would be minor. We need further studies to scrutinize the effects of age on the white matter anatomy and its relationship with transformation methods. Another[SM27] important source of errors could be differences in imaging parameters, especially B0-susceptibility distortion. For example, images from 3T scanners are often more distorted than those from 1.5T scanners. After linear normalization, imperfect template-subject matching due to non-linear differences caused by individual anatomical differences, age-dependent differences, and image parameter-dependent differences remain as error sources, which leads to the imperfect correlation between the manual and automated approaches, as shown in Fig. 6. Our future efforts will focus on reducing these error sources by using higher-order, non-linear transformation, an age-matched template (if necessary), and imaging methods with less distortion.