Whole-heart 3DCSPAMM allows for fast and reliable functional imaging of the beating heart by mapping the deformation of magnetization patterns throughout the left ventricle and cardiac cycle [1]. Its clinical use has been limited mainly due to (i) the time-consuming post-processing of tagged MR images, and (ii) derivation of clinically relevant biomarkers from the data. Regarding the post-processing of tagged MR images, multiple approaches have been proposed [2]; however, consensus about a preferred algorithm is still lacking. Regarding clinically relevant prognostic information, biomechanical models are increasingly considered as a potential tool to characterize patient condition and predict patient outcome through the integration of patient-specific data including ventricular strain [3]–[5]; however, the integration of finite element biomechanical models and image data remains challenging.

The objective of the present work was to develop, validate and analyze a finite element digital image correlation approach to extract ventricular strain data from MR images, which can be applied to both 3DCSPAMM images and conventional multi-slice cine images.

Whole-heart 3DCSPAMM and standard multi-slice balanced SSFP cine data were acquired in healthy volunteers on a clinical 1.5T scanner (Philips Achieva, Best, The Netherlands). Imaging parameters for 3DCSPAMM were: spatial resolution 3.5×7.7×7.7mm3 reconstructed to 1×1×1mm3, temporal resolution 28ms, tagging distance 7mm; geometrical stack alignment was performed by incorporating navigator offsets and rigid image registration. Imaging parameters for cine were: spatial resolution 1.2x1.2x8mm3, temporal resolution 70ms. Both 3DCSPAMM and cine images were interpolated in time using a Lanczos filter in MeVisLab (MeVis Medical Solutions AG, Bremen, Germany) to yield ca. 50 frames/cycles. Cine images were spatially interpolated as well, still using MeVisLab Lanczos filter, in order to obtain ca. isotropic voxel sizes. Upon manually segmenting the left ventricle on 3DCSPAMM and cine images using MeVisLab, reference left ventricular finite element meshes at end-diastole were generated using GMSH [6] (Figure 1).

The finite element digital image correlation approach developed here is based on the hyperelastic warping FEBio [7] plugin, which implements the method described in [8], [9]. Briefly, considering a moving image $$$M$$$, a target image $$$T$$$, and a triangulation of an object on the moving image $$$\Omega$$$, the method consists of finding the displacement field $$$\underline{U}$$$ that minimizes the functional $$J\left(\underline{U}\right) = \int_\Omega \frac{k}{2} \left(M\left(\underline{X}\right)-T\left(\underline{X}+\underline{U}\right)\right)^2 d\Omega + \int_\Omega \psi \left(\underline{U}\right)\left(\underline{X}\right)d\Omega$$, where $$$k$$$ denotes a penalty coefficient (the first term being a similarity term), and $$$\psi$$$ the hyperelastic strain energy potential (so that the mechanical energy acts as a regularization). In the present work, a simple Neo-Hookean potential with unit stiffness is assumed, so that $$$k$$$ is the only parameter of the method. Successive images are registered and the successive displacement fields are combined so as to map the motion of the reference mesh.

Optimal penalty factors were determined for cine and 3DCSPAMM images by computing peak circumferential, longitudinal and circumferential-longitudinal strains over a wide range of penalties, and noting the penalty that corresponds to a converged average strain and minimal standard deviation. The strain computation was then validated against a validated implantation of the SinMod method [10] within 3DTagTrack (Gyortools LLC, Zurich, Switzerland). Furthermore, strain data was obtained from cine and 3DCSPAMM images, and compared. Finally, the sensitivity of the method to image/mesh misregistration was investigated by manually shifting the reference mesh before applying the warping algorithm.

Optimal penalty factors were identified as $$$k=16$$$ for cine images, and $$$k=0.5$$$ for 3DCSPAMM (Figure 2).

The proposed method provided similar circumferential strain data compared to the validated SinMod method (Figure 3). Reduced variation of circumferential strain indicates more reliable tracking of the proposed method relative to SinMod.

Global circumferential and longitudinal strain components extracted from cine were found to be similar to strain data extracted from 3DCSPAMM images; however, the circumferential-longitudinal strain component, i.e., the ventricular twist, was not captured appropriately from cine images due to the lack of contrast within the myocardial wall (Figure 4).

Strain extraction from 3DCSPAMM images was found to be much less sensitive to misregistration compared to cine (Figure 5), due to the smoothing of the boundaries in 3DCSPAMM images.

Finite element digital image correlation is a promising approach to derive strain data from whole-heart 3D tagging data. In contrast to strain mapping from cine images, image correlation on 3D tagging data captures ventricular twist in agreement with physiological values, and is less sensitive to image misregistration. Accordingly, the method may be more appropriate compared to cine-based strain analysis as reported previously [11], especially when nonhomogeneous strain fields are expected, for instance in infarcted hearts, and for which intra-ventricular wall contrast is necessary.