Advanced Analysis of High-Resolution Cardiac Diffusion Tensor Images

About This Grant

Diffusion tensor MRI (DTI) can resolve distinct microstructural patterns in the myocardium but is challenging to perform in the heart because of its motion. We, and others, have used an approach in which the diffusion gradients null the first and second moments of motion (velocity and acceleration), which allows the diffusion of water to be resolved during free breathing. The principal challenge of this approach, however, is the presence of artifacts at air-tissue interfaces, such as the lateral wall of the heart. Consequently, we have recently developed an approach where the DTI acquisition is focused on an inner volume centered on the intraventricular septum. This approach allows the spatial resolution of the acquired data to be increased by almost 10-fold and eliminates image artifacts. We have leveraged the increased quality and resolution of these images to introduce a new analysis approach for cardiac DTI, which we term DTI phenomapping. While previous approaches have analyzed DTI-derived metrics on per-segment or per-patient scale, our approach allows the data to be quantified in each voxel, individually. With this approach, each voxel can be described by an array of 10 DTI-derived parameters, which can be combined to form a subject or population matrix. This matrix can then be analyzed with techniques such as principal component analysis and hierarchical clustering. We have shown, using this approach, that the myocardium in healthy individuals contains 4-5 distinct microstructural clusters and that the myocardium in well- compensated aortic stenosis maintains its gross microstructural coherence. While this approach represents a paradigm shift in analysis of cardiac DTI data, several challenges remain. DTI data are inherently signal-to-noise constrained, which we address using a specifically designed 64-channel cardiac coil. However, the most basal slices in the heart still cannot be analyzed due to insufficient signal. In addition, our analysis is confined to the septum, which can be limiting if the pathology of interest does not involve the myocardium diffusely. We hypothesize here that machine learning can be used to address some of these technical limitations and allow cardiac DTI to predict adverse cardiac events in patients with severe aortic stenosis. In addition to the scientific goals of this K25 application, the award will provide a mechanism for the applicant to transition the focus of his research from material science to medical imaging and machine learning and to obtain further training and mentorship in these fields. In Aim 1 of the proposal, we will focus on techniques to extend the anatomical coverage of the DTI phenomapping approach to more segments of myocardium. In Aim 2, we will refine the signal processing techniques used in the phenomapping approach and use unsupervised machine learning to identify new patterns. In Aim 3, we will use the improved approach to longitudinally follow subjects with severe aortic stenosis. These scientific goals will be combined with a dedicated program in scientific and professional growth, including formal coursework, training in biostatistics, and the responsible conduct of research.