Abstract
Accurate tracking of myocardial motion in 3D echocardiography (4DE) is essential for assessing cardiac strain and identifying pathologies. However, low signal-to-noise ratios in 4DE images often lead to physically implausible motion estimates. We present a semi-supervised deep learning framework that utilizes a Multi-Layered Perceptron (MLP) as an intelligent regularizer. By incorporating biomechanical constraints—such as incompressibility and periodicity—the model bridges the domain gap between synthetic training data and real-world clinical data. Our results demonstrate significant improvements in motion tracking correlation and the ability to accurately localize myocardial infarcts.
Introduction
Tracking the movement of the heart muscle (myocardium) using 3D+time echocardiography (4DE) is a critical goal for non-invasive cardiology. Measuring strain—the degree of muscle deformation—allows for the early detection of heart failure and permanent injury, such as infarction. Despite its potential, 4DE is plagued by high noise levels, causing existing tracking algorithms to produce "jittery" or physically impossible motion estimates.
While deep learning offers a potential solution for data smoothing, the lack of "ground truth" labels for in vivo cardiac motion presents a significant challenge. Models trained on synthetic simulations often fail to generalize to the noisy reality of clinical ultrasound scans, a classic domain adaptation problem.
Methods
The proposed framework employs a Multi-Layered Perceptron (MLP) network designed to act as an "intelligent regularizer." Unlike standard Convolutional Neural Networks (CNNs), the MLP is applied to localized 4D patches of pre-extracted motion data, where spatial filters are less critical.
The network takes noisy initial displacement estimates from standard tracking methods (e.g., block matching or non-rigid registration) as input. To bridge the gap between synthetic and real data, the authors introduced a semi-supervised loss function. While synthetic data utilizes a supervised loss against known ground truths, in vivo data is governed by biomechanical constraints:
- Incompressibility: Myocardial tissue volume must remain roughly constant (divergence-free constraint).
- Periodicity: The heart should return to its initial state after a complete cardiac cycle.
Furthermore, the framework supports multi-view learning, allowing it to process inputs from multiple tracking algorithms simultaneously and learn a superior "consensus" motion field.
Key Insights and Ideas
A central finding of this research is that biomechanical constraints are essential for effective domain adaptation. Without these constraints, the network tends to learn the "identity function," simply returning the noisy input data.
As shown in t-SNE visualizations, unconstrained networks clearly separate synthetic data from real-world data. The introduction of physics-based loss functions forces the network to find a shared latent representation, effectively teaching the model to recognize physiologically plausible heart beats across both domains.
Key Results
The proposed method was validated using both synthetic datasets and in vivo canine studies (acute ischemia and chronic infarct models). Key findings include:
- Improved Correlation: On real-world data, the MLP regularizer significantly improved the Pearson correlation with sonomicrometer ground truth. For instance, a Radio-Frequency Block Matching tracker's correlation improved from 0.01 to 0.26.
- Multi-View Performance: Combining multiple tracking inputs allowed the model to achieve a correlation of 0.67, outperforming any individual method.
- Infarct Localization: The model’s predicted strain-based injury maps matched manually traced postmortem infarct regions with a DICE score of 0.65, demonstrating clinical utility for identifying damaged tissue.
Conclusion
This work demonstrates that physics-based regularization is a powerful tool for domain adaptation in medical imaging. By enforcing biomechanical rules, deep learning models can learn from synthetic data and successfully generalize to noisy, unlabeled clinical environments. Future research will focus on expanding the diversity of synthetic training data and incorporating higher-resolution ground truth from modalities like MRI tagging.
Summary updated