Non-invasive Activation Times Estimation Using 3D Echocardiography

Adityo Prakosa, Maxime Sermesant, Hervé Delingette, Eric Saloux, Pascal Allain, Pascal Cathier, Patrick Etyngier, Nicolas Villain, Nicholas AyacheSTACOM/CESC 2010 · DOI

TL;DR

This paper presents a novel machine learning approach to non-invasively estimate cardiac activation times from 3D echocardiography. By learning the relationship between heart motion (kinematic descriptors) and electrical activity from a simulated electromechanical model, they can predict activation patterns in real patients, which is crucial for understanding and treating heart conditions like heart failure and optimizing cardiac resynchronization therapy.

What Problem Does This Paper Address?

Understanding the intricate "mechano-electrical coupling" within the heart – how the heart's mechanical motion relates to its electrical activity – is fundamental for diagnosing and treating various cardiac conditions. Current methods for mapping electrical activity, like intracardiac electrophysiological mapping, are invasive. While non-invasive techniques like electrocardiographic imaging exist, they are not yet widely validated or available. This gap leaves a critical need for non-invasive methods to assess cardiac electrical propagation, especially for guiding therapies like Cardiac Resynchronization Therapy (CRT), where precise lead placement and timing are vital for patient outcomes. The authors aim to bridge this gap by leveraging readily available 3D echocardiography data.

Methods

The core of this work lies in a machine learning approach, specifically Kernel Ridge Regression, to learn the complex relationship between cardiac motion and electrical activation.

  1. Motion Estimation: They first estimate cardiac motion (displacement and strain) from 3D echocardiography sequences using an incompressible diffeomorphic demons algorithm. This algorithm is an advanced image registration technique that accounts for the near-incompressibility of the myocardium.
  2. Electromechanical Model for Training: Since direct simultaneous acquisition of both motion and electrical data from the same patient is challenging, they create a synthetic training dataset using an electromechanical model of the heart. This model simulates various cardiac conditions (normal sinus rhythm, Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), and biventricular pacing) to generate corresponding kinematic (motion) and electrical (contraction force) data.
  3. Kernel Ridge Regression: This learned relationship is then used to predict activation times from the kinematic descriptors extracted from real patient echocardiography data. They use a Radial Basis Function (RBF) kernel for the regression, allowing them to model non-linear relationships.
  4. Parameter Optimization: Crucially, they optimize the hyperparameters (λ, σ) of the Kernel Ridge Regression model using a leave-one-out cross-validation approach, minimizing the Predicted Residual Sum of Squares (PRESS) statistic.

Datasets

  • Training Data: Synthetic data generated from an electromechanical model of the heart. This model was used to simulate four cardiac cases: normal sinus rhythm, LBBB, RBBB, and biventricular pacing.
  • Evaluation Data:
    • Synthetic data from the electromechanical model (cases 1-3 for training, case 4 for testing).
  • Real 3D echocardiography data from a patient undergoing CRT, acquired in two modes: sinus rhythm and left ventricle pacing.

Key Insights and Ideas

  • Bridging the Mechano-Electrical Gap: The paper demonstrates that it's possible to infer electrical activation patterns from mechanical motion data, even without direct electrical measurements. This opens doors for non-invasive cardiac electrical mapping.
  • The Power of Simulation for Training: Using a sophisticated electromechanical model to generate training data is a clever workaround for the difficulty of acquiring simultaneous, high-fidelity motion and electrical data from real patients.
  • Strain as a Stronger Predictor: The results suggest that strain measurements are more correlated with activation times than displacement measurements, implying that the deformation of the heart muscle provides more direct information about the underlying electrical events.
  • Generalization Capability: The Kernel Ridge Regression model, trained on a subset of simulated data, showed good generalization capabilities when applied to predict activation on unseen data points and even on a patient case.

Key Results

  • Simulated Data: The method achieved a Root Mean Squared Error (RMSE) of 0.0016 MPa between predicted and ground truth contraction force on simulated data. The RMSE for activation time prediction was 7 ms for the first three simulated cases and 37 ms for the fourth (unseen) case.
  • Patient Data: Qualitative results on patient data showed predicted contraction force curves that, while noisy, captured the expected global bell shape. The bull's-eye plots of estimated activation times corresponded to expected patterns, showing earlier activation in the left ventricle during pacing, as anticipated.

Limitations and Future Work

  • Noise in Patient Data: The estimated contraction forces from patient data were found to be noisy, and the curves sometimes exhibited negative parts, which is not physically expected for contraction forces.
  • Discrepancy between Simulated and Real Motion: A potential factor for the observed discrepancies is the difference between simulated strains/displacements and those estimated from real patient data via non-linear registration.
  • Reference Image Sensitivity: Errors in selecting the reference end-diastolic image can lead to significant errors in strain estimation.
  • Patient Anatomy Sensitivity: The current model was evaluated on a patient whose anatomy was used to build the electromechanical model. Further evaluation on more patients is needed to assess the method's sensitivity to individual anatomy.

Future Work: * Test the method on larger patient cohorts with diverse stimulation protocols. * Investigate the sensitivity of the approach to strain estimation accuracy and reference image selection. * Explore learning directly from simultaneous intracardiac electrophysiological mapping and 3D echocardiography data from the same patient, once such data becomes available.

Summary updated

BibTeX
@inproceedings{prakosa2010non,
  author = {Adityo Prakosa and Maxime Sermesant and Hervé Delingette and Eric Saloux and Pascal Allain and Pascal Cathier and Patrick Etyngier and Nicolas Villain and Nicholas Ayache},
  title = {{Non-invasive Activation Times Estimation Using 3D Echocardiography}},
  booktitle = {STACOM/CESC},
  year = {2010},
  doi = {10.1007/978-3-642-15835-3\_22},
}
Figure