Continuous Deformation Modeling via Implicit Neural Fields

27 concepts21 papers1 proposal

Whether tracking the motion of a beating heart, simulating the drape of a fabric, or aligning misaligned medical scans, we often need to model how space deforms. Historically, this meant dividing space into discrete pieces: voxel grids for volumetric medical images, or meshes with vertices for physical simulations. To find the displacement at a point between grid nodes, we relied on discrete interpolation methods like trilinear or B-spline interpolation.

However, discrete representations scale poorly with resolution and introduce blocky artifacts or high-frequency noise. A powerful alternative has emerged: representing deformations as a continuous function using coordinate-based neural networks, often called Implicit Neural Representations (INRs) or Neural Fields. Instead of storing displacement vectors on a grid, we train a multi-layer perceptron (MLP) to map a continuous spatial coordinate directly to its corresponding displacement or velocity vector:

$$f_\theta(\mathbf{x}) = \mathbf{u}$$

This paradigm shift fundamentally changes how we compute, regularize, and scale deformation fields across medical imaging, computer graphics, and physical simulations.


Resolution Independence and Grid-Free Representation

In traditional discrete representations, the resolution of a deformation field is locked at the time of creation. If you need finer details, you must increase the grid density, which causes memory requirements to explode.

By parameterizing the deformation map as a neural field, the representation becomes entirely discretization-agnostic. As highlighted by Chen et al. in their survey of medical image registration, INRs allow transformations to be evaluated at arbitrary spatial coordinates and varying sampling densities. This eliminates the blocky artifacts typically associated with discrete data structures when mapping high-dimensional spaces.

For instance, in physical simulations, Kairanda et al. introduced a cloth simulator that replaces discrete mesh vertices with a continuous neural deformation field. This allows the geometry to be queried at any resolution, maintaining consistent details without being tied to a fixed mesh topology. Similarly, Chen et al. applied INRs to the Material Point Method, demonstrating that a continuous deformation map enables adaptive resolution and super-resolution during simulations. In medical imaging, Tian et al. and Sun et al. showed that this grid-free formulation allows for highly accurate, continuous image registration without the constraints of traditional spatial partitioning.

Analytical Gradients via Automatic Differentiation

In deformation modeling, we rarely care only about the displacement itself; we almost always need its derivatives. For example, to ensure a medical image registration is biologically plausible (diffeomorphic), we must penalize configurations where the deformation folds over itself. This requires computing the Jacobian matrix of the deformation field. In physical simulations, calculating strain and stress requires spatial derivatives of the displacement map.

With a discrete grid, these derivatives must be approximated using finite differences, which are sensitive to grid resolution and introduce numerical errors.

With INRs, the deformation map is an analytically defined, differentiable neural network. As El Hadramy et al. demonstrate, this allows for the exact computation of complex, higher-order gradients through automatic differentiation rather than discrete approximations. Spears et al. leverage this property for susceptibility distortion correction in MRI, using the smooth, analytically tractable derivatives of the INR to compute precise geometric corrections. Similarly, Chen et al. bypass explicit discretization entirely by analytically computing deformation gradients and velocities directly through network differentiation.

The Smoothness Inductive Bias

Deformations in the real world—whether the squishing of soft tissue or the movement of a fluid—are typically smooth and continuous. When training a model to predict these transformations, we want a representation that naturally favors smooth solutions over noisy, discontinuous ones.

Coordinate-based MLPs possess an inherent inductive bias toward smoothness. Ignatiev et al. exploit this characteristic by using an INR as a geometric prior for continuous deformations, showing that it naturally avoids the "shattered gradient" problem common in high-resolution warping.

This smoothness is particularly valuable when modeling velocity fields to generate diffeomorphic transformations. Instead of predicting displacements directly, methods like those proposed by Han et al. use an MLP to parameterize a continuous velocity field, which is then integrated over time. Because the underlying neural network is inherently smooth across the spatial domain, the resulting deformation maps preserve anatomical topology and prevent tearing or overlapping, even when capturing complex anatomical structures.

Bypassing the Cubic Memory Wall

In 3D applications, voxel-based methods suffer from a severe limitation: cubic memory scaling ($O(N^3)$). Doubling the resolution of a 3D grid requires eight times more memory, quickly hitting hardware limits for high-resolution volumetric data.

INRs break this scaling law. Because the deformation field is parameterized by the weights of a neural network rather than explicit grid nodes, the memory footprint depends on the capacity of the network, not the resolution of the evaluation grid. Lowes et al. demonstrated this advantage when tracking the left ventricle myocardium across a cardiac cycle, showing that INRs offer a highly memory-efficient alternative that avoids the prohibitive memory costs of traditional convolutional architectures. Tian et al. similarly showed that querying displacements at arbitrary coordinates via an MLP requires up to five times less memory than traditional grid-based deep learning approaches while maintaining state-of-the-art accuracy.

Dynamic and Hyperparameter Conditioning

A major strength of the neural field formulation is its extensibility. Because the network takes coordinates as inputs, we can easily append other continuous variables to the input vector, allowing the deformation field to be conditioned on dynamic contexts or hyperparameters.

  • Temporal Conditioning: Li et al. extended spatial INRs into the spatiotemporal domain by feeding both spatial coordinates and time variables $(x, y, z, t)$ into the network. This enables continuous 4D deformable image registration and seamless frame interpolation.
  • Hyperparameter Conditioning: Deformable registration requires balancing image alignment with regularization constraints. Finding the right balance usually requires hours of retraining for different hyperparameter values. To solve this, El Hadramy et al. and El Hadramy et al. conditioned the INR directly on the regularization hyperparameter itself. At inference time, users can dynamically tune the smoothness of the deformation field in real time without retraining the network.

Key Foundational Work

An efficient data-driven tissue deformation model (2009) — contributed 1 concept elements to this direction.
Learning to Importance Sample in Primary Sample Space (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_displacement_field_modeling_ce1f3754,
  title = {continuous displacement field modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-displacement-field-modeling/rdr_ce1f37547a1044ef8ac56ba13046bfb1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_displacement_field_modeling_ce1f3754_v1,
  title = {continuous displacement field modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-displacement-field-modeling/rdr_ce1f37547a1044ef8ac56ba13046bfb1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake