Recent advances in cardiac diffusion tensor imaging (DTI) have enabled robust imaging of the in-vivo human heart [1,2]. This allows to non-invasively assess myofiber and myosheet orientations, which has been a long standing bottleneck of patient-specific cardiac modeling [3]. However, in view of the relatively low scan efficiency and scan time constraints in-vivo, cardiac coverage and signal-to-noise ratio are limited. To this end, data extrapolation [4] and noise reduction [5] techniques have been proposed. The objectives of the present work are to propose an approach to (i) extract the most significant features from noisy myofiber and myosheet angle maps measured with in-vivo DTI, and (ii) extrapolate the data across the entire left-ventricle based on a limited number of acquired short-axis views.