1.4. Methodology
1.4.1. Cardiac Motion Features Extraction
1.4.1.2. MR Tagging Processing Methodology
In order to get relevant biomarkers, it is mandatory to design a robust methodology for the MR-T processing and the subsequent tensor calculation. We have made use of
the previously mentioned WHARP method to extract the phase images. In addition, we have extended this methodology for the computation of the 3D deformation gradient tensor using SA and LA images on the intersection of the different slices (for points on which LA images were not available, 2D motion has been reconstructed).
This processing pipeline for MR-T was first proposed inCordero-Grande et al.(2011) so as to obtain locally smooth estimates of the image phase. Subsequently, the tensor es- timates have been refined using the LAD method rather the common LS approach, less robust towards outliers and artifacts. In Cordero-Grande et al. (2016) this method has been extended to allow the introduction of an overdetermined set of tagged images (with different stripe orientations), so that artifact presence in the HARP images is greatly diminished. The method has been compared to its non-windowed counterpart in the es- timation of the ST, showing great improvement in terms of accuracy and reproducibility both with and without multiple orientations. However, the processing pipeline requires a fine tuning in some stages, especially for the window and filter designs. We have extended the WHARP procedure so as to become completely automatic and adaptive both in the spatial and the spectral domains. We have resorted to an angled-steered analysis window, whereas the band-pass filter has been designed to narrow in the modulation direction and to widen in the orthogonal direction. No parameters are manually set since their values are partially based on the information available at the Digital Imaging and Communi- cation in Medicine (DICOM) headers and in additional information directly estimated from the data. Window and filter designs have been thoroughly described and analyzed in Chapter 5. All the updates have been merged in the proposed pipeline, which we explain in detail in Section 1.4.1.3.
On the other side, we have also checked the ability of the methodology to provide motion descriptors with clinical utility. First, we have proposed an image processing met- hodology to distinguish fibrotic tissue by assessing the local mechanical properties of the myocardium. The analysis of the local deformation patterns has been carried out with the purpose of finding an agreement between hyperenhanced zones in late enhancement images and areas in the myocardium with abnormal tensor values. The agreement is mea- sured taking as Ground Truth (GT) the manual scar segmentation carried out by two cardiologists in the late enhancement images (Sanz-Est´ebanez et al., 2015).
We have also made use of the statistical learning theory for the diagnosis of a variety of cardiomyopathies states using MRI-derived features. We have proposed a (sequential) two-stage classification scheme capable of distinguishing between heterogeneous groups of HCM and healthy volunteers. Results have shown that well-established classifying metho- dologies can be arranged to accommodate the study of HCM with acceptable performance using the aforementioned tensorial measurements, even for reduced and unbalanced sam- ple sets. The sequential classification methodology and the design of each of its inner stages have been described in more detail in Chapter 6.
Although strain measurements have proven useful for cardiomyopathy screening, they cannot tell apart the different genotypes behind these diseases. Therefore, the prognosis in the appearance of fibrotic tissue cannot be achieved with these descriptors. For these
purposes, rotation measures have implemented so as to provide additional information on myocardial mechanics as a complement of standard pump function indices. However, most of the rotation parameters described in the literature assume a cylindrical geometry for the LV, which implicitly requires a fixed RA. Therefore, as stated in Section1.3.4any misalignment will induce estimation errors and thus will greatly hinder the subsequent diagnosis and classification. To alleviate this, we have introduced a novel local rotation descriptor based on robust tensorial measurements that relates the presence of increased vorticity values with the fibrotic tissue in the heart. Rotation is estimated by means of the curl operator without any influence of global myocardial parameters, such as RA or cavity radius. With such a design, we can carry out a regional comparative study in patients with different forms of LV hypertrophy coming from different etiologies, namely, HCM and secondary forms of LV hypertrophy, as well as healthy subjects.
In Chapter 2, which constitutes the first core paper of this Thesis, we have included this study which relates the presence of fibrosis with local vortices in myocardial tissue. To the best of our knowledge, this is the first work that makes use of the curl operator for the prognosis of fibrotic tissue.
Furthermore, we have also inspected the design of more appropriate tag patterns, within a single acquisition scenario, by exploring the use of multiple peaks in the k-space, as opposed to multiple orientations. We have assessed, by means of a computational phantom, optimal tag orientations and spacings of the stripe pattern by minimizing the Frobenius Norm Difference (FND) between the GT tensor and the estimated material deformation gradient tensor. In addition, we have measured performance loss with respect to multiple orientations in a real setting. Results indicate that, for a single acquisition, multiple peaks, as opposed to multiple orientations, are indeed preferable. Chapter 7will explain in more detail these ideas.