ILPU: Iterative Laplace-Based Phase Unwrapping via Bi-Level Optimization
Michael Fedders, Jakob Schattenfroh, Yanglei Wu, Noah Jaitner, Tom Meyer, Jakob Jordan · 2026
Read ILPU: Iterative Laplace-Based Phase Unwrapping via Bi-Level Optimization by Michael Fedders, Jakob Schattenfroh, Yanglei Wu, Noah Jaitner, Tom Meyer, Jakob Jordan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-31460.txt.
Introduction
Fedders and five colleagues present ILPU, an iterative method for unwrapping phase in magnetic resonance imaging. Phase-based MRI serves susceptibility mapping, field mapping, thermometry, and elastography, but the measured phase wraps around, and unwrapping it in noisy regions is error-prone. A wrong unwrap corrupts every map built on top.
ILPU splits the job into two levels. A lower solver recovers a smooth phase increment from an incremental Poisson equation by a discrete cosine transform. An upper solver refines an integer offset map with quality-guided smoothing and a restricted local search. Adaptive penalties weighted by quality keep low signal regions stable. Tests on 2D and 3D brain phase images against hand-unwrapped references judge the method where standard Laplace unwrapping struggles.
Catalog record
| CREATOR | Michael Fedders, Jakob Schattenfroh, Yanglei Wu, Noah Jaitner, Tom Meyer, Jakob Jordan |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.31460
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ILPU: Iterative Laplace-Based Phase Unwrapping via Bi-Level Optimization Thanks: Funding by German Research Foundation (DFG) projects 513752256 FOR5628, CRC1340, GRK2260 BIOQIC, CRC 1540 EBM, GU 172614-1, German Federation of Industrial Research Associations (AIF) project KK5611902 BM4, and Medical Scientist Pilot Program of the Charité is greatly appreciated. These sponsors had no role in the study design, collection, analysis and interpretation of data, writing of this manuscript or the decision to submit the article for publication. Thanks: All the authors are with Department of Radiology, Charité – Universitatsmedizin Berlin, 10117, Berlin, Germany (email:[email protected]) Michael Fedders, Jakob Schattenfroh, Yanglei Wu, Noah Jaitner, Tom Meyer, Jakob Jordan Abstract Phase unwrapping is an essential preprocessing step for phase-based MRI applications including susceptibility mapping, field mapping, thermometry, and MR elastography. We present Iterative Laplace-Based Phase Unwrapping (ILPU), a bi-level optimization algorithm in which a lower-level solver recovers a continuous phase increment from an incremental Poisson equation via the discrete cosine transform (DCT), while an upper-level solver refines an integer offset map through quality-guided spatial regularization and a restricted local search. This coupling yields robust unwrapping in low-SNR regions through adaptive smoothness penalties and quality-weighted regularization. We evaluated ILPU on 2D and 3D brain MRI phase images against manually unwrapped reference data, using standard Laplace unwrapping, Flynn, and SEGUE as comparison methods. In 2D, ILPU achieves accuracy comparable to SEGUE. In 3D, ILPU attains a relative error of 2.12% versus 67.59% for SEGUE and 81.02% for Laplace, demonstrating a clear advantage in volumetric unwrapping. The algorithm operates at O ( N log N ) O(N\log N) complexity per iteration through DCT-based Laplacian estimation, and is numerically faster than both Flynn and SEGUE while preserving superior accuracy. These results indicate that the bi-level optimization framework provides a robust and computationally efficient solution for phase unwrapping in MRI. Keywords: Phase unwrapping, Bi-level optimization, Laplacian unwrapping, Magnetic resonance imaging, Elastography. 1 Introduction Phase unwrapping recovers the continuous phase field from wrapped measurements confined to the principal value range, serving as an essential preprocessing step for phase-based magnetic resonance imaging (MRI) applications. The complex MRI signal encodes tissue properties in both magnitude and phase, but the arctangent operation used to extract the phases inherently wraps values to (-\pi,\pi] , creating discontinuities where the true phase exceeds these bounds. Accurate phase unwrapping enables numerous clinical applications including quantitative susceptibility mapping (QSM) for tissue characterization and iron quantification in neurodegenerative diseases [ 21 ] , proton resonance frequency shift-based thermometry for monitoring thermal ablation procedures [ 30 ] , B 0 field mapping for geometric distortion correction in echo-planar imaging [ 18 ] , phase-contrast flow imaging for cardiovascular assessment [ 26 ] , and MR elastography (MRE) for tissue mechanical property characterization [ 28 , 31 , 32 ] . MRE encodes tissue displacement from externally applied mechanical vibrations in the phase images, where unwrapping errors propagate directly into stiffness estimates, making phase quality critical across a broad range of clinical applications, including liver fibrosis staging and tumor characterization. The challenges are particularly important in MRE due to low displacement amplitudes in stiff tissues, discontinuities at organ boundaries, and the need for robust unwrapping across multiple vibration frequencies and motion encoding directions [ 24 ] . Despite decades of research, robust phase unwrapping in MRI remains challenging due to several fundamental difficulties. Phase measurements in low signal-to-noise ratio (SNR) regions become unreliable, with noise amplified through the arctangent function generating residues, i.e., pixels where closed-loop path integration yields non-zero multiples of 2\pi [ 33 ] . These residues violate the integrability conditions required for unique unwrapping solutions and can propagate errors throughout the image. Near air-tissue interfaces such as paranasal sinuses and auditory canals, large magnetic susceptibility differences create steep phase gradients that often exceed the Nyquist sampling criterion, resulting in spatial aliasing manifested as multiple 2\pi wraps between adjacent pixels [ 36 ] . Existing phase unwrapping algorithms can be broadly categorized into local and global methods. Local methods rely on local consistency and integration paths. They typically start from high-quality seed points or phase-coherent regions and propagate the unwrapping decision to neighboring pixels or voxels. Because the decision is made locally and then propagated, these methods are usually fast and effective when a reliable quality map is available, but they are sensitive to the starting seed, path selection, and local error propagation. This family includes classical path-following and quality-guided methods such as those of Goldstein et al. [ 15 ] and Herráez et al. [ 16 ] , as well as regional methods such as PRELUDE and SEGUE (Speedy rEgion-Growing algorithm for Unwrapping Estimated phase) [ 20 ] . PRELUDE, a gold standard among region-growing methods, groups voxels into phase-coherent regions and merges them by integer offsets. SEGUE builds on PRELUDE and addresses its speed limitation, and serves as one of the reference methods in this work. Global methods enforce consistency across the entire image or volume, typically through energy minimization or global constraint satisfaction, and can be broadly categorized as either discrete or continuous. Discrete global methods optimize integer offsets or discontinuities directly. This includes network flow methods by Costantini [ 7 ] and Chen and Zebker [ 4 , 5 ] , as well as statistical formulations based on Bayesian inference and Markov random fields [ 29 , 37 ] . Continuous global methods estimate the phase field by minimizing a continuous objective over the full image. This includes the minimum L^{p} -norm formulation by Ghiglia and Romero [ 13 ] , weighted least-squares methods, and Laplacian or Poisson-based approaches [ 12 , 38 , 34 , 23 ] . Laplacian approaches exploit the fact that, under appropriate regularity conditions, the Laplacian of the unwrapped phase can be inferred directly from the wrapped observations. This enables a fast recovery of the phase through solution of a Poisson equation with suitable boundary conditions, often with discrete cosine transform (DCT) solvers and O(N\log N) complexity [ 38 ] . Weighted variants incorporate quality maps by assigning reliability-based weights to phase gradients, but pure Laplacian approaches still lack explicit mechanisms to resolve integer offset ambiguities that may persist in noisy regions or in the presence of rapid phase variations. The Flynn method [ 10 ] also belongs to this group. It formulates unwrapping as a global optimization problem that minimizes weighted discontinuities, using quality maps to concentrate unavoidable errors in low-reliability regions. Although this method is computationally less convenient for large volumetric datasets, Flynn remains an accurate and well-established reference for 2D data. For this reason, we include it as the other reference throughout this work. Bi-level optimization provides a natural framework for hierarchical problems where one optimization task (the upper level or leader) constrains another (the lower level or follower) [ 6 ] . This structure arises naturally in inverse problems where lower-level reconstruction parameters are optimized by upper-level criteria and vice versa. For phase unwrapping, bi-level optimization enables natural decomposition. This paper presents ILPU (Iterative Laplace-Based Phase Unwrapping), a bi-level optimization algorithm for MRI phase unwrapping. The lower level employs DCT-based Laplacian unwrapping to estimate regularized phase increments with O(N\log N) complexity, while the upper level refines integer-valued offset maps through quality-guided spatial regularization using