TLDR: A new method using Implicit Neural Representations (INRs) has been developed for automatically quantifying heart muscle motion and strain from MRI scans. This AI model predicts continuous left ventricular displacement with high accuracy (2.14 mm RMSE) and low strain errors, outperforming existing deep learning methods. Crucially, it’s also significantly faster (380x faster than the most accurate baseline), making it suitable for large-scale analysis of cardiac function.
Assessing the function of the heart is crucial for diagnosing and managing cardiovascular diseases. One key indicator is myocardial strain, which measures the relative deformation of the heart muscle. Early detection of left ventricular (LV) dysfunction through strain analysis can predict adverse outcomes.
Current methods for analyzing heart motion, such as Cardiac Magnetic Resonance (CMR) feature tracking, often fall short by only considering tissue boundaries, making regional strain measurements unreliable. CMR tagging, which places a grid on the heart muscle to track its movement, is more accurate but traditionally requires time-consuming manual analysis.
While deep learning has been applied to automate tag tracking, these methods have their own limitations. They can suffer from low output resolution, be affected by image artifacts like tag fading, or be slow during the analysis process.
A Novel Approach with Implicit Neural Representations
Researchers have introduced a new method that uses Implicit Neural Representations (INRs) to overcome these challenges. This innovative approach predicts continuous left ventricular displacement without needing complex optimization during the analysis phase. INRs are essentially neural networks that learn continuous functions from discrete data, allowing for highly detailed and arbitrarily resolvable displacement functions.
The proposed INR model is designed to learn a realistic and continuous displacement function from sparse tracking data. It also incorporates a physiological constraint: myocardial incompressibility, meaning the heart muscle’s volume should not change significantly. This helps ensure the predictions are biologically plausible.
At its core, the INR method uses a multi-layer perceptron (MLP) that learns the displacement function. This MLP is “conditioned” on unique latent codes derived from the image frames by a convolutional neural network (CNN) encoder. These latent codes act as a compact representation of the image pair, allowing the model to generalize across different cases. Modulation networks then adaptively adjust the MLP’s behavior based on these latent codes.
The training process involves optimizing all these components together using a combination of three loss functions: a position loss to ensure accurate tracking of points, a latent code loss to prevent overly large latent code magnitudes, and a Jacobian loss to enforce the incompressibility of the myocardium.
Also Read:
- CardioComposer: Guiding 3D Anatomy Generation with Precision and Flexibility
- Improving Heart Condition Diagnosis Through Physiology-Aware AI
Impressive Performance and Speed
The new INR method was rigorously evaluated on 452 test cases from the UK Biobank dataset and compared against three existing deep learning baselines: BioTag, SynthTag, and DeepTag. The results were highly promising.
The INR method achieved the best tracking accuracy with a root mean squared error (RMSE) of 2.14 mm. It also demonstrated the lowest combined error in global circumferential (2.86%) and radial (6.42%) strain. Furthermore, this method is remarkably fast, being approximately 380 times faster than the most accurate baseline, DeepTag, processing 1,250 slices per second compared to DeepTag’s 250 slices per second.
These findings highlight that INR-based models are highly suitable for accurate and scalable analysis of myocardial strain in large CMR datasets. The method’s ability to predict smooth, physiologically consistent motion, even in the presence of image artifacts, marks a significant advancement in the field.
While the current study primarily focused on healthy volunteers and relied on manually tracked points (which have some inherent imprecision), the researchers plan to explore synthetic data to reduce the need for manual tracking and validate the method on clinical datasets with various cardiovascular diseases. For more details, you can refer to the full research paper.
In conclusion, this generalized implicit neural representation offers an efficient and accurate way to track intramyocardial motion from CMR tagging, promising better assessment of cardiac function for improved diagnosis and management of heart conditions.
