Cardiac motion tracking drops manual atlas alignment: CardioDrift beats every baseline on UK Biobank and M&M
Synopsis
Cardiac motion tracking from cine magnetic resonance previously either lacked anatomical correspondence across subjects or relied on a slow, error-prone manual atlas fit. CardioDrift generalises coherent point drift to jointly register the left-ventricular endocardium, epicardium and right-ventricular atlas surfaces, and estimates per-slice translations inside the same probabilistic model to correct slice misalignment, fitting and propagating the atlas automatically from short-axis segmentation contours alone. On 1,301 UK Biobank subjects and the M&M dataset, its rigid fitting and non-rigid propagation lead every distance and overlap metric against the strongest baseline.
Interpretation
Cardiac motion can now be represented directly on a 3D atlas mesh whose points correspond anatomically across subjects, using only short-axis cine CMR segmentation contours as input. Image-domain methods represented motion in each subject's own coordinate space and established no anatomical correspondence across individuals, while atlas-mesh methods needed a semi-automatic step to fit the atlas to the subject space. On 1,301 randomly selected UK Biobank subjects, the rigid stage reached Dice coefficients of 90.72%, 70.25% and 81.01% for the left ventricle, myocardium and right ventricle, above all compared methods.
Atlas fitting and frame-by-frame propagation sit in one probabilistic registration, with the endocardial, epicardial and right-ventricular surfaces aligned to their anatomically corresponding contours under a single shared transformation. Standard coherent point drift neither enforces correspondence between each atlas surface and its anatomical counterpart nor corrects slice misalignment, which limits fitting accuracy. The non-rigid stage reached the lowest HD95 and mean surface distance over the full UK Biobank cycle, raising myocardial Dice from 64.91% for coherent point drift, 60.41% for CPD-Net and 60.90% for DeepMesh to 82.11%.
Slice-shift correction is embedded in the same expectation-maximisation iterations of the rigid stage, estimating in-plane translations for each short-axis slice. Without correction the fitted atlas is visibly offset in a subset of subjects, and that error propagates into the non-rigid stage. On UK Biobank, enabling correction raised myocardial Dice from 55.76% to 70.25% and improved every left-ventricular and myocardial metric, while the three right-ventricular metrics were marginally worse; on M&M, seven of nine metrics improved.
Perspective
The demonstrated scope is the full cardiac cycle for 1,301 UK Biobank subjects and expert annotations at end-diastole and end-systole for 134 M&M subjects; the input is contour points extracted from short-axis cine CMR segmentation maps, with the three atlas surfaces decimated to 359, 459 and 402 points.
Slice correction helps less on M&M than on UK Biobank because residual misalignment from multi-protocol acquisition falls outside a single in-plane translation per slice; M&M reference masks exist only at end-diastole and end-systole, so temporal consistency over intermediate frames is tested on UK Biobank alone; the equations must be solved independently at each frame, and the formulation has no term controlling temporal smoothness of the output mesh sequence.