Abstract
Standard spline interpolation techniques typically assume a uniform, homogeneous domain, which can lead to significant artifacts and "blurring" when signals cross structural boundaries. This paper presents Domain-Informed Spline Interpolation (DIBSI), a novel framework that integrates supplementary structural information directly into the interpolation basis. By constructing shift-variant kernels that adapt to local domain characteristics, DIBSI achieves superior signal reconstruction accuracy in inhomogeneous environments, such as the human brain.
Introduction
In signal processing, interpolation is a fundamental tool for upsampling and reconstruction. Traditional methods, like B-splines, rely on shift-invariant kernels, effectively treating the underlying space as a blank slate. However, many real-world signals are physically constrained by complex domains. In neuroimaging, for instance, functional brain activity is intrinsically linked to specific tissue types (e.g., gray matter vs. white matter). Standard interpolation often "mixes" signals across these biological boundaries, leading to potential misinterpretations of functional connectivity.
The Homogeneity Gap
The core limitation of standard spline interpolation is the assumption of domain homogeneity. When upsampling fMRI data, a standard cubic B-spline kernel assumes every voxel belongs to the same type of space. Because these kernels are shift-invariant—meaning they maintain the same shape regardless of their position—they cannot respect the sharp transitions between different tissue types. The proposed Domain-Informed Spline Interpolation (DIBSI) addresses this by replacing static kernels with context-aware, shift-variant basis functions.
Methodology: Constructing a Domain-Informed Basis
DIBSI redefines the interpolation problem by incorporating a set of subdomain functions that characterize the environment. These functions typically represent probability maps of different tissue types or regions.
The framework constructs a tailored kernel $\phi_k(x)$ for each sample point $k$. This kernel is composed of a dominant component, representing the primary subdomain at that location, and a residual component that adjusts for neighboring subdomains. A key feature is the inclusion of a tunability parameter, $\gamma$, which allows users to control the "aggressiveness" of the adaptation to domain boundaries.
As shown in Figure 2, DIBSI's basis functions (rows f and k) adapt their shape based on the underlying domain functions (row a), whereas standard B-splines (rows d and i) remain uniform. Despite this shift-variance, the authors demonstrate that the DIBSI basis maintains critical mathematical properties, including stability (Riesz basis) and the partition of unity.
Experimental Results
The authors validated DIBSI using both synthetic simulations and real-world neuroimaging data from the Human Connectome Project (HCP).
Synthetic Validation
Monte Carlo simulations on 10,000 synthetic signals showed that DIBSI consistently outperformed standard B-spline interpolation (BSI) across all tested spline orders. The benefit of DIBSI was particularly pronounced for higher-order splines, where standard methods tend to "smear" signals more due to their wider support.
Figure 4 illustrates how DIBSI (red line) more accurately tracks the underlying synthetic signal (black line) compared to standard BSI (blue line), especially at the sharp transitions defined by the domain functions $d_1(x)$ and $d_2(x)$.
Neuroimaging Application
When applied to fMRI data, DIBSI produced functional maps that more accurately respected anatomical structures like sulci (brain grooves). While standard BSI blurred activity across these gaps, DIBSI maintained the separation, providing a more biologically plausible reconstruction of brain activity.
Conclusion
Domain-Informed Spline Interpolation offers a significant advancement for signal processing in structured environments. By moving away from shift-invariant assumptions, DIBSI provides sharper, more accurate reconstructions that respect known physical boundaries. While the method introduces higher computational costs due to the need to calculate unique kernels for each point, its superior performance in high-stakes fields like medical imaging makes it a compelling alternative to traditional splines.
Summary updated