Capítulo V: Análisis de la Industria
5.5. Matriz de Perfil Competitivo MPC
Tract-based spatial statistics (TBSS)
Tract-based spatial statistics (TBSS) [12] is a method that aims to improve the two main limita- tions of VBS analyses of FA: proper spatial normalization, and sensitivity to amount of spatial smoothing. On the one hand, the spatial normalization employed by VBS does not guarantee a perfect alignment of white matter fibers. For example, it has been shown that VBS studies may report residual registration misalignments as FA reductions, particularly when there are variations in ventricular sizes [68]. On the other hand, it has been shown that VBS-FA results depend strongly on the amount of smoothing [69].
Briefly, TBSS has four steps: (i) registration of all subjects’ FA images to a common tem- plate using non-linear registration. At this stage perfect alignment is not expected; (ii) creation of a FA average image and extraction of a skeleton by removing non-maximal values, per- pendicularly to the local structure. Areas of low FA and/or high inter-subject variability are removed also; (iii) projection of each subject’s aligned FA image onto the skeleton; and (iv) voxel-wise statistics across subjects on the skeletonized versions of the respective FA maps. Using the same method, it is also possible to project white matter MD maps onto the skeleton for statistical analysis.
TBSS studies of temporal lobe epilepsy have shown general FA reduction in the corpus callosum, thalamus, internal/external capsule, temporooccipital and fronto temporal connec- tions [70]. Changes in the uncinate fasciculi are more prevalent in MTS patients compared to non-MTS patients [44]. MD increases have been also been detected with TBSS in clusters of temporal white matter predominantly ipsilateral to seizure onset [39, 70].
TBSS results have been compared to those obtained from VBS analyses showing that TBSS may be more sensitive to white matter changes than VBS studies [39]. However it is important
deterministic tractography probabilistic tractography
Figure 1.15: Tractography. Tractography constructs global map of white matter connectivity, either deterministically or probabilistically. In probabilistic tractography maps of confidence in connection to theseedsare obtained, here shown as isosurfaces.16
to consider the fact that the hippocampal formation does not form a part of the FA skeleton, and thus VBS could provide complimentary information for mesial structures. In the same manner, TBSS results have been combined with VBM analysis of gray matter, showing a correlation between hippocampal gray matter volume and FA white matter [44].
White matter tractography and tractometry
While MD and FA are scalar values characterizing the diffusion process (Figure 1.12), the eigenvectors of the diffusion tensor can be used to identify coherent anatomical patterns of dif- fusion, more specifically, to identify white matter fibers where diffusion is highly anisotropic. This technique is known as white mattertractography and allows the identification of white matter pathologies such as trauma, tumours and inflammation.
Tractography studies can use either deterministic or probabilistic methods. Determinis- tic tractography relies solely on information from the source DTI study, such as the main eigenvector and the FA values on every voxel to guide streamlining algorithms. In contrast, probabilistic methods use white matter atlases [71] providing a map of relative probability that a voxel belongs to a particular fiber given its location, DTI properties, and the similarity be-
15Adapted from Brain and Language 131 (2014), J. Campbell and B. Pike,Potential and limitations of diffusion
tween its orientation and that of the voxel’s atlas [72]. Both generally requireseed voxels to initialize the estimation of tracts. Seed selection can be manual, or semi-automatic by means of a registered pre-seeded template [43]. Small FA values in white matter can misguide the estimation of tracts. Therefore, it is common to restrict these algorithms to FA regions above a predetermined threshold (between 0.15 and 0.3) [43, 72, 73].
Tractography holds promise for identifying white matter connections in-vivo. Neverthe- less, current techniques are limited by many factors including: (i) sensitivity to initialization parameters, including seed location; (ii) limited spatial resolution that undermines the iden- tification of sub-voxel structures and the resolution of fiber crossing paths; (iii) the lack of a gold standard to validate results. Additionally, results often need to be manually adjusted by an expert to control for bad or misidentified tracts.
Tractometry is the process of assigning quantitative measures to the reconstructed stream- lines from tractography [74]. In studies of temporal lobe epilepsy, both the average FA and MD are be quantified per tract [43, 73]. Though most TLE tractometry studies agree on reporting FA loss in tracts ipsilateral to the seizure onset, a more detailed interpretation of results is chal- lenging, since the extent and number of regions vary from study to study and these are highly dependent on the specific tracing algorithm employed. For example, two independent research groups reported that patients with L-TLE present more widespread white matter abnormalities than right TLE (R-TLE) patients [72, 75], while a third group reported the opposite [43].
Imaging analyses show pieces of a much larger puzzle in which findings are heterogeneous and highly dependent of individual patient pathologies. This poses the question: how can these clues be combined to obtain a diagnosis? Machine learning algorithms contribute to this goal by discovering distinctive patterns in clinical data that can be diagnostically useful.