Introduction Reconstructing each cardiac sub-structures such as left ventricle (LV), left atrium(LA), aorta, right ventricle (RV), right atrium(RA), pulmonary artery (PA) has been studied for extensively purposes such as hemodynamic analysis and percutaneous and transcatheter therapeutic treatment planning. Reconstructing complex shapes such as crescent RV shape, LV/LA/aorta and RV/RA/PA junction is not a trivial task especially in cardiac magnetic resonance (CMR) images with high anisotropy (large slice spacing).
Methods: Cine MR images were acquired from ten healthy volunteers. Contours of LV, LA, arota, RV, RA, and PA were delineated. We use left cardiac structure for example. To reconstruct the structure in between one LV contour and two LA/AO contours, we utilized the following extra steps. (a) The convex hull of LA and AO was computed: ConvHull; (b) The alignment between ConvHull and LV was computed via cyclicDTW; (c) The alignment between points from LA and AO yet not included in ConvHull was computed via DTW; (d) The voronoi diagram of LA and AO was computed constrained to LV. The voronoi edge between LA region and AO region was annotated as LA-AO-Edge; (e) Results from (a-c) constitutes the bounding convex hull of expected LV/LA/AO junction. We modified all alignment line between LA and AO from (c) to parabolic curves whose apex were determined by LA-AO-Edges from (d). The modification led to a saddle shape in LV/LA/AO junction which is our current best knowledge.
Results: The average of processing time is about 4.9s on a 2.5 GHz CPU desktop. The overlapping ratio between the reconstructed model and three chamber view has an average 0.92 Dice Index.
Conclusions: We proposed a cyclicDTW based method to tackle both one-to-one contours alignment as well as one-to-multiple. Establishing the alignment between short axis contours is equivalent to reconstructing the cardiac shape from sparse cross sectional contours.