🤖 AI

Learning Cardiac Motion Priors for Implicit Neural Representations

This paper evaluates four strategies for learning cardiac motion priors to enhance Implicit Neural Representations (INRs), demonstrating that while all methods improve adaptation over random initialization, meta-learning achieves the best overall performance in tracking accuracy and adaptation trajectory.

Original authors: Andrew Bell, George Webber, Andrew P King, Steffen E Petersen, Muhummad Sohaib Nazir, Alistair Young

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrew Bell, George Webber, Andrew P King, Steffen E Petersen, Muhummad Sohaib Nazir, Alistair Young

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Mapping a Moving Heart

Imagine you are trying to draw a map of how a heart muscle moves and beats. The heart is constantly changing shape, squeezing, and relaxing. In the medical world, we use special MRI scans (called "tagged" images) to see this movement.

The researchers are using a type of artificial intelligence called an Implicit Neural Representation (INR). Think of an INR not as a standard computer program, but as a super-smart, flexible clay sculpture.

How it works: You tell the clay, "Here is the heart at time 1," and "Here is the heart at time 2." The clay reshapes itself to match the movement perfectly.
The Problem: If you start with a brand new, shapeless lump of clay for every single patient, it takes a very long time to sculpt it correctly. Also, if you push the clay too hard in the wrong direction at the start, it might get stuck in a weird shape that doesn't look like a real heart.

The Solution: Learning "Muscle Memory"

The paper asks: Can we teach the clay some "muscle memory" before we even start sculpting a specific patient? Instead of starting with a blank lump, can we start with a lump that already looks a bit like a beating heart?

The researchers tested four different ways to give the AI this "muscle memory" (called priors):

The "Group Average" (Population Prior): Imagine taking 100 different heart sculptures, squishing them all together, and making one giant "average" heart. You use this average as your starting point.
- Result: It's a good start, but because it's an average, it might be a bit "mushy" or compromise too much. After a while, it gets stuck and doesn't improve much.
The "Class Consensus" (Consensus Prior): Imagine 100 students each sculpting a heart perfectly on their own. Then, you take the final shape of every student's sculpture and blend them together to make one "best guess" starting point.
- Result: This worked very well and was simple. It was a strong starting point that kept getting better over time.
The "Personalized Kit" (Auto-Decoder): Imagine a master sculptor who has a toolbox of different "tools" (latent codes). For every new patient, the AI picks the right tool from the box to customize the starting clay.
- Result: This was the fastest at handling big, dramatic movements early on. It could quickly grab a large deformation and fix it. However, it hit a "ceiling" and stopped improving after about 10 tries.
The "Learning to Learn" (Meta-Learning): Imagine training a sculptor not just to make a heart, but to learn how to learn how to make a heart. The AI is taught the best way to start and the best way to adjust its steps as it goes.
- Result: This was the overall winner. It started strong and, unlike the others, kept getting better and better even after 50 attempts. It found the smoothest, most accurate path.

The Experiment

The team tested these methods using real heart scan data from the UK Biobank (a massive database of health information). They measured:

Accuracy: How close the AI's movement map was to the actual human heart movement.
Speed: How quickly the AI could figure out the movement.
Physics: Did the heart look like it was breaking or folding in impossible ways? (The AI was programmed to ensure the heart stayed "incompressible," like real muscle).

What They Found

Starting with a "blank slate" (Random Initialization) was bad. It was slow and often inaccurate.
All four "muscle memory" methods were much better than starting from scratch.
The "Auto-Decoder" was the sprinter: It fixed big errors very quickly in the first few seconds.
The "Meta-Learning" method was the marathon runner: It had the best long-term performance, continuing to refine the heart's movement smoothly over a longer period.
The "Consensus" method was the reliable all-rounder: It was simple, effective, and kept improving steadily.

The Takeaway

The paper concludes that how you start matters. You don't just need a smart AI; you need to give it a smart starting point.

If you need a quick fix for a big movement, use the Auto-Decoder.
If you want the most accurate, long-term result, use Meta-Learning.
If you want a simple, effective solution, the Consensus Prior (averaging weights) is a great choice.

Essentially, the researchers proved that teaching the AI "how a heart usually moves" before asking it to solve a specific case makes the whole process faster, more accurate, and more reliable.

Technical Summary: Learning Cardiac Motion Priors for Implicit Neural Representations

Problem Statement

Implicit Neural Representations (INRs) offer continuous, compact representations of cardiac motion fields, making them well-suited for motion estimation. However, fitting an INR to a specific image sequence is an ill-posed inverse problem that is computationally expensive and highly sensitive to the optimization trajectory. While learned priors have been proposed to guide optimization toward physiologically plausible solutions and accelerate adaptation, the specific application of learned priors to cardiac motion INRs remains under-explored. Existing approaches often rely on simple prior formulations or lack systematic analysis of the resulting motion fields and adaptation trajectories.

Methodology

The authors propose a framework to learn and compare four distinct strategies for constructing cardiac motion priors for INRs. The core motion model is a Neural Velocity Field (NVF), an INR that maps spatial coordinates and time to instantaneous velocity vectors.

1. Motion Representation and Objective

Neural Velocity Field (NVF): The cardiac motion is represented as a velocity field $v_\theta: \mathbb{R}^{d+1} \to \mathbb{R}^d$ . Lagrangian motion trajectories are obtained by integrating the ordinary differential equation defined by the NVF using a second-order Runge-Kutta method.
Periodic Encoding: To enforce cardiac periodicity, time inputs are mapped to periodic encodings using sine and cosine functions.
Loss Function: The optimization minimizes a weighted sum of four terms:
- Similarity Loss ( $L_{sim}$ ): Normalized Cross-Correlation (NCC) between the current frame and forward/backward warped adjacent frames.
- Smoothness Loss ( $L_{smooth}$ ): Penalizes spatial and temporal gradients of the velocity field.
- Volume Preservation ( $L_{volume}$ ): Encourages incompressibility by penalizing deviations of the deformation gradient determinant from 1.
- Topology Preservation ( $L_{fold}$ ): Prevents folding by penalizing negative determinants of the deformation gradient.

2. Prior Formulations

The study compares four strategies for initializing the NVF parameters ( $\theta$ ):

Population Prior: A single INR is jointly optimized across all training instances to find a global average solution ( $\theta_P$ ).
Consensus Prior: INRs are independently optimized for each training case, and the resulting parameters are averaged to form a prior ( $\theta_C$ ).
Auto-decoder Prior: Each training instance is assigned a unique latent code $z_i$ . A shared decoder network with modulation parameters $\kappa$ is jointly optimized with the latent codes. During adaptation to a new case, the decoder is fixed, and a new latent code is optimized.
Meta-learned Prior: Utilizing second-order Meta-SGD, the method learns an optimal initialization $\theta_0$ and layer-wise adaptation step sizes $\gamma$ . The objective is to minimize the loss after a fixed number of inner-loop adaptation steps.

3. Experimental Setup

Dataset: Short-axis tagged Cardiac Magnetic Resonance (CMR) images from the UK Biobank (500 training, 100 validation, 100 test slices).
Architecture: All priors initialize a SIREN network (3 hidden layers, 256 nodes).
Evaluation: Performance is measured by tracking accuracy (mean Euclidean distance of reference points), segmentation overlap (Dice score), and mechanical plausibility (Jacobian determinants and strain). Adaptation is evaluated over 5 steps (early performance) and 50 steps (long-term trajectory).

Key Results

Overall Improvement: All learned priors substantially outperformed random initialization. Mean displacement error was reduced from 3.45 mm (random) to a range of 1.66–1.85 mm after five adaptation steps.
Meta-learning: Achieved the lowest overall displacement error (1.66 mm) and maintained the best adaptation trajectory over 50 iterations, continuing to improve well beyond the 5-step optimization horizon used during training.
Auto-decoders: Demonstrated the fastest recovery of large deformations, achieving the lowest end-systolic (ES) displacement error (2.11 mm) and largest strain magnitudes near end-systole. However, performance plateaued after approximately 10 iterations.
Consensus vs. Population: The simple consensus prior was effective and continued to improve over 50 iterations. In contrast, the population prior, while providing a strong starting point, began to deteriorate after 30–40 iterations, likely converging to a compromise solution that hinders further adaptation.
Latent Dimensionality: Varying the latent dimension in auto-decoders (from 32 to 1024) had little effect on the adaptation trajectory, suggesting the plateau is due to the modulation mechanism rather than latent space capacity.

Significance and Claims

The paper claims to provide the first systematic evaluation of prior learning strategies for INR-based cardiac motion estimation. The authors assert that:

Adaptation Speed and Trajectory: The choice of learned prior dictates both the speed of initial adaptation and the long-term optimization trajectory.
Trade-offs: Different priors offer different trade-offs; auto-decoders are superior for rapid recovery of large deformations, while meta-learning offers the most robust long-term optimization.
Feasibility: Learned priors make the computationally expensive per-case optimization of INRs more viable by enabling faster convergence and more physiologically plausible motion fields.

The work concludes that while simple heuristics like weight averaging (consensus) are effective, more principled approaches like meta-learning and auto-decoders provide distinct advantages depending on the specific requirements of the motion estimation task (e.g., speed of convergence vs. long-term accuracy).