In this paper, we tested the atlas-based quantification tool for the anatomical evaluation of CP patients. Typically, with the large extent of anatomical alteration in the CP population, the detection of anatomical abnormality by MRI is often straightforward and, thus, the aid of sophisticated image analyses may not be needed if only abnormality detection is the concern. What is not straightforward; however, is the quantification of the visible abnormalities, which could be an important resource for understanding the pathology, classifying patients, studying the anatomy-functional correlation, and possibly in the future, predicting functional outcomes and designing an intervention at the early phase of brain development.
In the quest to establish correlation between anatomical damages and functional outcomes, several attempts have been made in the past (Glenn et al., 2007
; Hoon and Melhem 2000
; Huppi and Dubois 2006
; Kanda et al., 2004
; Ludeman et al., 2008
; Maenpaa et al., 2003
; Nagae et al., 2007
; Rose et al., 2007
). One of the major targets of these past studies had been to identify anatomically important areas, the injury of which could be a surrogate marker of functional impairment. By focusing on several hypothetically important areas, these approaches, however, had a limited power for functional correlation. For example, even for a well-established relationship between motor function and the corticospinal tract, the correlation in CP patients is far from clear, and, in fact, there is evidence that injuries in the sensory pathway may also be a significant factor (Hoon Jr et al., 2009
). We apparently need a more comprehensive and quantitative approach to describe the anatomical status of the patients. The proposed atlas-based quantification would provide means by which to quantitatively analyze the status of multiple areas at once. This is an essential first step toward comprehensive anatomy-function correlation studies of the CP population.
In this study, we introduced several new information and tools to further investigate the anatomical description of the pathology. First, the use of DTI allows us to delineate many substructures within the white matter, which are not visible with conventional MRI. Using this contrast, there is the potential that we can characterize injuries in specific white matter structures sensitively (Neil et al., 2002
; Son et al., 2007
). Second, we introduced 3D quantitative analyses based on the atlas-based segmentation of the entire brain into 110 areas. By measuring the volume, FA, ADC, and axial and radial diffusivity, we created a 5 × 110 matrix to profile the anatomical status of each patient quantitatively.
Accuracy of the brain segmentation and power analysis
The results of the whole-brain segmentation, based on the WMPM, were compared with manual segmentation (Appendix Table 2
). The Kappa value, Jaccard similarity metrics, and Dice coefficients showed that the segmentation accuracy rivals that of inter-rater variability in all measured structures except for the corpus callosum, for which the inter-rater accuracy was significantly better than the automated method, but all measurements showed a high level of accuracy (Kappa > 0.80). The atlas-based segmentation showed noticeably lower accuracy for the cingulum, which was expected for this type of long and narrow structure; even a one-pixel shift along the entire length would lead to substantially lower accuracy. High inter-rater variability (large standard deviation) and accuracy (low Kappa) were observed for the superior corona radiata (SCR). This is understandable because there is no clear boundary of the corona radiata to separate it from the anterior, superior, and posterior portions. There is the potential that the WMPM-based segmentation could introduce objective criteria to delineate such structures.
The specificity was high (>99%) and the false negative rate was low (<2%) for all the regions, in all the comparisons. However, the overall sensitivity (82%) was slightly lower and the false negative rate (18%) was slightly higher for the automated segmentation compared to that achieved with inter (86% and 14%) and intra-raters (88% and 12%). Nevertheless, the correlation between automated measurements and visual scoring for atrophy was as strong as the correlation between visual scorings performed by different evaluators, indicating that the automated method is a practical alternative for the labor intense and time consuming visual semi-quantitative analysis.
The required level of accuracy depends on the expected effect size and the normal range of anatomical variability. The preliminary power analysis data, based on the 13 normal subjects is, thus, important for judging the usefulness of this approach. The variability observed for each of the 5 × 110 anatomical matrices is influenced by the true anatomical variability and accuracy of the measurements. indicates the amount of abnormality needed in each variable to detect a difference between groups using an unpaired t-test, with an alpha less than 0.05. It also indicates the number of subjects needed per group to detect a 10% change in each variable in the same conditions, with a power of 0.95. We need, for example, larger samples or a larger degree of abnormalities to detect volumetric differences than is needed to detect FA or ADC differences. In the same way, the sample size or the differences in subcortical gray matter must be larger than that for structures in the white matter to detect the same degree of contrast difference. On the other hand, smaller differences or a smaller sample size is needed to detect volume differences in the subcortical gray matter compared to the white matter.
Quantitative anatomical profiling of each patient
Based on the atlas-based segmentation tool, we performed quantitative whole-brain evaluation of the anatomical status of each CP patient, as shown in . In agreement with the results from the power analysis, smaller variations in FA, compared to volume, can be detected as beyond the normal limits (predicted values ± 2 standard deviations - SD). In this example, a decrease of 12% in the left superior longitudinal fasciculus FA places the subject under the normal curve −2SD, while a decrease of 45% in volume is needed to have the same effect. Regardless of the relative insensitivity of the volume measurements, the volume z-score maps of the CP patients detected as many abnormal regions as the FA z-score maps, indicating the large amount of brain atrophy in this population. Once segmented, we can create an extensive report that will capture the characteristic anatomic features of each patient. For diseases like CP, which is known as a highly inhomogeneous entity, this type of evaluation would provide a highly individualized view of brain anatomic status. The proposed approach may provide information to identify not only anomalous regions, compared to normal subjects, but also their correlation with clinical performance. It is potentially possible to identify regions that determine subgroups in terms of anatomical and/or functional characteristics. In CP, this ‘global’ approach of analyzing the combined effect of multiple regions, tracts, or systems is particularly promising since clinical evidence points to multiple abnormalities that, together, determine the clinical status. In this paper, our results are descriptive and are limited to quantitative presentation of the abnormalities. For future investigations, this type of quantitative analysis will be a key to the effective utilization of MRI data acquired at an early phase of brain development to enable better prediction of prognosis.