Continuous 3D Myocardial Motion Tracking via Echocardiography

TL;DR

This work tackles the challenge of accurate 3D myocardial motion tracking from echocardiography by introducing NeuralCMF, a self-supervised implicit neural representation that jointly models the 3D cardiac structure and 6D motion across the entire cardiac cycle. The method utilizes dual SIREN streams (static and dynamic) and a physics-informed loss with imaging, motion, cycle, and regularization terms to supervise motion estimation without paired ground-truth data, and it accommodates both 2D and 3D echocardiography inputs. Across STRAUS and multi-view 2DE/3DE datasets, NeuralCMF delivers state-of-the-art tracking accuracy, enables CHD classification from LV motion with high accuracy, and demonstrates robust 3D reconstruction and motion querying beyond traditional methods. The approach holds practical significance for clinical workflows, enabling precise, device-agnostic, and bedside-ready motion analysis with potentially broader applicability to other organs and imaging modalities.

Abstract

Myocardial motion tracking stands as an essential clinical tool in the prevention and detection of cardiovascular diseases (CVDs), the foremost cause of death globally. However, current techniques suffer from incomplete and inaccurate motion estimation of the myocardium in both spatial and temporal dimensions, hindering the early identification of myocardial dysfunction. To address these challenges, this paper introduces the Neural Cardiac Motion Field (NeuralCMF). NeuralCMF leverages implicit neural representation (INR) to model the 3D structure and the comprehensive 6D forward/backward motion of the heart. This method surpasses pixel-wise limitations by offering the capability to continuously query the precise shape and motion of the myocardium at any specific point throughout the cardiac cycle, enhancing the detailed analysis of cardiac dynamics beyond traditional speckle tracking. Notably, NeuralCMF operates without the need for paired datasets, and its optimization is self-supervised through the physics knowledge priors in both space and time dimensions, ensuring compatibility with both 2D and 3D echocardiogram video inputs. Experimental validations across three representative datasets support the robustness and innovative nature of the NeuralCMF, marking significant advantages over existing state-of-the-art methods in cardiac imaging and motion tracking.

Figures (11)

• Figure 1: We introduce a self-supervised method for estimating 3D motion trajectories throughout the cardiac cycle at any given point all at once. The figure illustrates the motion estimation of the 3D heart from end-diastole (ED) to end-systole (ES), representing the heart's contraction phase. The estimated motion trajectories are displayed on the right with each point's trajectory depicted in a different color for clarity. Though we have chosen to show only sparse trajectories of the left ventricular (LV) myocardium, our method has the ability to compute motion for all points across the entire 3D heart.
• Figure 2: The figure illustrates the workflow of our method. (a) Inputs consist of 4D space-time coordinates $(x,y,z,t)$. Corresponding outputs include the intensity value ${O}(x,y,z,t)$ at specified space-time locations, as well as forward and backward 3D motion vectors ${\vec{M}}=(\vec{m}_{t \rightarrow t+1},\vec{m}_{t \rightarrow t-1})$ that represent the movements of $(x,y,z)$ at successive time frames $t+1$ and $t-1$, respectively. (b) Our approach can be applied to both 2DE and 3DE videos. (c) The methodology utilizes two main optimization strategies: physical motion consistency loss, ensuring that intensity remains constant within both the heart's tissues and blood-filled areas throughout the entire cardiac cycle, and physical cardiac cycle consistency loss, which aligns the process with the physiological fact that the heart muscle returns to its initial shape upon completing one full cardiac cycle.
• Figure 3: This figure presents the qualitative results of point tracking using our method, VoxelMorph balakrishnan2019voxelmorph, and Co-AttentionSTN ahn2023co on the STRAUS alessandrini2015pipeline, 2DE, and 3DE video datasets. The query points that need to be tracked during the ED phase are displayed in the first column. The following three columns visually demonstrate the movement of these points across the full cardiac cycle. In the 2DE video datasets, only our results are showcased since other methods are incompatible with multi-view 2DE datasets.
• Figure 4: This figure illustrates the tracking performance of our method, VoxelMorph, and Co-AttentionSTN for points on the left atrial (LA) myocardium within 3DE video datasets. The initial column highlights the reference points set for tracking during the ED phase. The subsequent three columns visually demonstrate the movement of these points across the full cardiac cycle. It is of significance to note that our method demonstrates superior accuracy in tracking points on the mitral valve annulus, a challenge not achieved by either VoxelMorph or Co-AttentionSTN.
• Figure 5: This experiment demonstrates that our model offers a more precise motion estimation, with the warped image aligning closely to the actual physiological structure of the heart. The figure illustrates a comparison of our warped image results with those obtained using other deep-learning-based methods. The warping process involves taking a 3D image at the ED phase and transforming it with estimated displacements to create corresponding 3D images at the ES phase. Subsequently, we slice the apical four-chamber view for a detailed comparison. For clarity, the LV myocardium is segmented, and the DICE index is computed between the ground truth and warped ES phase images.