In medical imaging, "registration" is the process of aligning two or more images—such as a pre-operative MRI and a post-operative CT—into a single coordinate system. While simple in concept, the deformation required to align biological tissues must be physically plausible. If a model simply predicts independent displacement vectors for every pixel, it risks "folding" the image, effectively creating overlapping tissues or tearing anatomical structures.
To solve this, researchers have converged on a framework that treats image registration not as a static displacement, but as a continuous fluid flow. By predicting a Stationary Velocity Field (SVF) and integrating it over time, models can guarantee diffeomorphic transformations: mappings that are smooth, invertible, and topology-preserving.
The Problem of Anatomical Folding
Traditional deep learning registration models often predict a dense displacement field $u$, where the transformed position of a voxel $x$ is simply $x + u(x)$. However, there is nothing in a standard neural network to prevent two different voxels from being mapped to the same location, or for the spatial ordering of voxels to be reversed. In a clinical context, this would mean a model "predicting" that a piece of the liver has disappeared or folded into itself.
As Peizhi Chen and Chin‐Ling Chen observe in their work on respiratory motion, these "physically impossible" deformations are a primary hurdle for medical AI. To maintain the integrity of the anatomy, the transformation must be a diffeomorphism.
Velocity Fields as the Engine of Deformation
Instead of predicting the final destination of each voxel, this research direction focuses on predicting the velocity at which voxels move. Imagine the image as a fluid. If we define a velocity field $v$, we can determine the path of any point by solving an Ordinary Differential Equation (ODE):
$$\frac{d\phi}{dt} = v(\phi(t))$$
where $\phi(0)$ is the identity mapping. By integrating this velocity field from $t=0$ to $t=1$, we arrive at a final deformation that is mathematically guaranteed to be invertible and smooth. James C. Gee and Tom Z. Jiahao formalize this in their NODEO framework, treating registration as a high-dimensional dynamical system where the network parameterizes the velocity field itself.
The Scaling and Squaring Algorithm
Integrating an ODE at every iteration of a deep learning model is computationally expensive. To make this practical for real-time clinical use, researchers like Huaqi Qiu and Chen Qin utilize the scaling and squaring method.
This algorithm exploits a property of stationary (time-independent) velocity fields: the total deformation can be computed by repeatedly composing small, incremental deformations. If we scale the velocity field by a factor of $1/2^N$ such that the movement is tiny, the deformation is approximately a simple translation. By "squaring" (composing the transformation with itself) $N$ times, we recover the full diffeomorphic mapping. This layer is differentiable, allowing models to be trained end-to-end while enforcing physical constraints.
From Voxel Grids to Neural Fields
While early work used Convolutional Neural Networks (CNNs) to predict velocity fields on a fixed grid, a newer wave of research uses Implicit Neural Representations. Instead of a grid, a coordinate-based MLP represents the velocity field as a continuous function.
- DNVF (Diffeomorphic Neural Velocity Fields): Kun Han and Shanlin Sun use MLPs with sinusoidal activations to represent continuous velocity fields. This allows for high-precision registration that is not limited by the resolution of a voxel grid.
- ST-NeRP: Liyue Shen and Lei Xing extend this to 4D (spatial-temporal) data, modeling how a patient's anatomy changes over time by predicting a velocity field that depends on both space and time coordinates.
Why It Works: The Inductive Bias of Flow
The success of this approach across diverse modalities—from breast MRI (Matthew French) to echocardiography (Zihao Wang)—stems from its strong inductive bias. By forcing the network to operate through a velocity-integration bottleneck, we prevent it from ever learning a non-physical solution.
Furthermore, using parameter-efficient representations like B-splines within the velocity field, as explored by Daniel Grzech and Ben Glocker, provides implicit smoothness. This reduces the degrees of freedom the model must learn, making it more robust when data is scarce or when quantifying uncertainty in the alignment.