Abstract
This paper introduces a new transport metric over the space of non-negative measures that interpolates between the quadratic Wasserstein ($W_2$) and the Fisher-Rao metrics. By generalizing the Benamou-Brenier dynamical formulation of optimal transport to include a source term in the continuity equation, the authors provide a framework for comparing measures with different total masses. The study proves the existence of geodesics, demonstrates that classical metrics are recovered as limiting cases, and provides a numerical scheme based on proximal splitting for applications in image interpolation.
Introduction
In computational anatomy, medical imaging, and computer vision, comparing two distributions of data—represented as measures—is a fundamental task. While Optimal Transport (OT), specifically the $W_2$ Wasserstein distance, has become a standard tool for this purpose, it suffers from a significant constraint: it requires the total mass of the two measures to be identical. In biological contexts, such as tracking the growth of a tumor or the development of brain tissue, mass is rarely conserved.
Conversely, the Fisher-Rao metric allows for changes in mass (intensity) but ignores the spatial geometry of the underlying space. This paper bridges these two paradigms by proposing the Wasserstein-Fisher-Rao (WFR) metric, a variational framework that simultaneously accounts for spatial transport and local mass variation.
The Dynamical Formulation
The authors derive the WFR metric by extending the Benamou-Brenier formulation of optimal transport. In the classical setting, mass $\rho$ moves according to a velocity field $v$ while obeying a strict continuity equation (mass conservation). The WFR framework introduces a source term $\zeta$ to this equation:
$$\partial_t \rho + \nabla \cdot (\rho v) = \zeta$$
Here, $\zeta$ represents the creation or destruction of mass at any point in space and time. The distance between two measures is then defined as the minimum of a convex functional that penalizes both the kinetic energy (the cost of moving mass) and the growth energy (the cost of changing mass intensity via the Fisher-Rao metric).
The Interpolation Parameter ($\delta$)
A key feature of the WFR metric is the parameter $\delta$, which defines the characteristic spatial scale of the interpolation:
- Large $\delta$: The cost of moving mass is low relative to the cost of creating it, causing the metric to behave like standard Optimal Transport.
- Small $\delta$: The cost of changing mass in place is low, causing the metric to behave like the Fisher-Rao metric.
Theoretical Insights
Existence of Geodesics
The authors rigorously prove that for any two non-negative measures, there exists an optimal path (a geodesic) that minimizes the proposed functional. This ensures that the metric is well-defined and that the "interpolation" between two images is physically and mathematically grounded.
The "Cut Locus" and the Security Distance
One of the most significant theoretical findings is the existence of a critical distance threshold, $\pi\delta$. This threshold determines the strategy the metric uses to transform one measure into another:
- If two mass points are separated by a distance less than $\pi\delta$, the metric prefers to physically move the mass from one location to the other (transport).
- If the distance exceeds $\pi\delta$, it becomes "cheaper" to let the mass at the source location fade out while simultaneously letting new mass fade in at the target location (growth/decay).
This behavior provides a natural way to handle objects that appear or disappear, preventing the "tearing" artifacts often seen in pure OT when it is forced to move mass across large, empty gaps.
Numerical Implementation
To solve the resulting convex variational problem, the authors employ a Douglas-Rachford proximal splitting algorithm. This approach is particularly effective for the staggered grids used in fluid dynamics. By formulating the problem on a space-time grid, the algorithm can efficiently find the global optimum, ensuring stability and avoiding the local minima that often plague non-convex image registration techniques.
Applications in Biological Imaging
The utility of the WFR metric is demonstrated through the interpolation of 2D segmented images of a developing brain. In this scenario, brain tissue grows non-uniformly, increasing the total mass over time.
Standard OT fails to produce a realistic interpolation because it attempts to conserve mass, often pulling tissue from one hemisphere to fill gaps in another. In contrast, the WFR metric allows for local growth. The resulting geodesics show a smooth, biologically plausible expansion where mass is added locally where needed, while still maintaining the spatial coherence of the brain's structure.
Conclusion and Future Work
The Wasserstein-Fisher-Rao metric provides a robust mathematical foundation for "unbalanced" optimal transport. By unifying spatial movement and mass change into a single convex optimization problem, it offers a powerful tool for image processing and shape analysis.
Future research directions include reducing the computational complexity to allow for real-time 3D MRI processing and investigating the metric's properties in the context of gradient flows for density evolution. While the choice of the parameter $\delta$ currently requires manual tuning based on the application scale, the WFR metric stands as a significant advancement in our ability to model dynamic, growing systems.
Summary updated