Source-linked AI summary
Primitive Representation Learning for Unsupervised Dynamic Contrast Enhanced MRI Reconstruction
Veronika Spieker, Wenqi Huang, Cemre Ariyurek, Liam Timms, Daniel Rueckert, Onur Afacan, Julia A. Schnabel, Sila Kurugol
TL;DR
Dynamic contrast-enhanced MRI needs accelerated reconstructions that preserve enhancement dynamics for reliable quantitative analysis. This paper introduces a primitive-based, disentangled reconstruction framework with separate anatomy, motion, and contrast tiers, achieving competitive reconstruction quality while preserving sharp aortic and renal enhancement.
Problem
DCE-MRI acceleration can compromise temporal signal fidelity, including the arterial enhancement dynamics needed for reliable quantitative parameter estimation.
Method
The framework models anatomy, motion-induced changes, and dynamic contrast enhancement as separate tiers with distinct temporal bases in a primitive-based representation.
Results
The method achieves reconstruction quality competitive with GRASP while preserving sharp aortic and renal enhancement that comparison methods over-smooth.
Takeaways & Limitations
Contrast primitives localize to expected enhancing anatomy and provide a geometrically interpretable separation of enhancement from static anatomy.
Takeaways & Limitations
Evaluation lacks a ground-truth reference, relies on GRASP comparison, and is restricted to a small in-vivo cohort.
Abstract
from arXiv · showhide
Reliable quantitative analysis of dynamic contrast-enhanced MRI requires high-quality spatiotemporal reconstructions at high undersampling rates. Scan-specific reconstructions using Gaussian and Gabor primitives have shown promising results without the need for large training datasets, but have not addressed the additional dimension of dynamic contrast. We propose a multi-dimensional, primitive based framework for dynamic contrast-enhanced MRI reconstruction that disentangles the underlying anatomy, the dynamic contrast enhancement, and residual motion into separate temporal basis functions, thereby enabling a geometrical interpretation of the representation. We show that this architecture achieves performance competitive with conventional reconstruction methods, both in reconstruction quality and in the accuracy of extracted aorta and kidney enhancement curves. The modular tier design extends naturally to additional dynamic factors and higher acceleration rates. Code available at https://github.com/compai-lab/2026-GaborDCE-spieker.
1 Introduction
DCE-MRI acceleration can compromise arterial input function dynamics and downstream quantitative estimation, while existing reconstructions may oversmooth temporal changes. The paper proposes an unsupervised, geometrically interpretable primitive-based framework that separately models anatomy, motion, and rapid contrast enhancement.
- Motivation: DCE-MRI acceleration must preserve arterial input function peak and shape for reliable quantitative estimation, but GRASP and XD-GRASP often oversmooth temporal dynamics.This trade-off is particularly consequential for applications such as MR urography for renal function assessment.
- Related work: Scan-specific unsupervised representations avoid large training datasets, which are difficult to obtain for abdominal MRI because motion-free, fully sampled references are challenging to acquire.However, coordinate-based implicit neural representations can overfit below resolution and provide limited interpretability, motivating explicit primitive representations.
- Related work: Gabor primitives improve primitive-based reconstruction by frequency modulation and low-rank temporal modeling, but this is insufficient for decoupled, nonsmooth contrast-enhancement dynamics.Contrast enhancement differs from motion because it is neither smooth nor periodic.
- Proposed method: The proposed DCE-MRI method adds a dedicated signal tier for rapidly arriving high-intensity contrast changes above the anatomical base signal, with separate temporal bases and regularization preserving sharp wash-in.The framework extends a motion-aware Gabor primitive reconstruction design to model distinct dynamic processes.
- Contributions: The paper presents the first primitive-based DCE-MRI reconstruction method and evaluates its reconstruction quality, quantitative parameters, and learned-representation geometry on in-vivo pediatric abdominal data.The multi-dimensional design and training strategy is intended to provide a geometrically interpretable, unsupervised representation.
2 Methods
The method represents dynamic MR images as sums of anisotropic Gabor primitives whose weights and geometry vary over time [10]. It decomposes each primitive’s signal intensity into anatomy, motion, and contrast-enhancement tiers with separate temporal bases.
- Primitive representation: Dynamic MR images are modeled as sums of N Gabor primitives with frame-varying weights and geometry [10].Each primitive is anisotropic and parameterized by position µ_n,t, scale s_n,t, rotation θ_n,t, and Gabor frequency ξ_n,t.
- Three-tier temporal model: Each primitive’s complex signal weight combines base anatomy, motion-coupled intensity changes, and dedicated dynamic contrast enhancement.The contrast tier captures enhancement after injection and its subsequent decay, while the motion tier captures motion-induced intensity changes.
- Three-tier temporal model: The anatomy, motion, and DCE tiers each use distinct temporal bases because their temporal behaviors differ.The original formulation used contrast and geometry bases for signal variation and cardiac motion [10]; the expanded model adds a dedicated DCE basis.
B Unsupervised Training A Spatiotemporal DCE Primitive Representation
The method represents dynamic contrast-enhanced MRI using separate contrast and geometry bases over primitive anatomy, motion, and enhancement factors. It fits these primitives through a density-compensated k-space objective with tier-specific temporal regularization and image-domain total variation.
- B Unsupervised Training A Spatiotemporal DCE Primitive Representation: Separate contrast and geometry bases model primitive intensity and shape/position, whose weighted rendered sum forms the DCE reconstruction and is transformed into k-space for unsupervised fitting.The contrast basis represents intensity, while the geometry basis represents primitive shape and position.
- B Unsupervised Training A Spatiotemporal DCE Primitive Representation: Primitive geometry combines static and temporally varying components for motion, while reusing the geometry basis ties enhancement and deformation to one temporal course.Geometry parameters include position, scale, rotation, and Gabor frequency.
- B Unsupervised Training A Spatiotemporal DCE Primitive Representation: The primitives are optimized with a density-compensated k-space objective that includes tier-specific temporal penalties and image-domain temporal total variation.The temporal penalties apply to the base, geometry, and contrast bases, while total variation penalizes frame-to-frame changes in rendered images.
- B Unsupervised Training A Spatiotemporal DCE Primitive Representation: Huber temporal regularization suppresses small temporal noise while preserving sharp bolus arrival, with stronger smoothing on base and geometry than contrast tiers.Each basis is normalized by its scale, and the contrast tier receives weaker regularization to keep enhancement sharp.
- B Unsupervised Training A Spatiotemporal DCE Primitive Representation: Joint optimization from scratch is unstable because the high-capacity base tier can absorb the contrast enhancement.The method therefore introduces staged optimization, although the supplied passage does not specify its individual stages.
3 Experimental Setup
The study evaluates dynamic MRI reconstruction on five pediatric acquisitions using a highly undersampled golden-angle 3D FLASH protocol. It combines qualitative assessment with similarity, temporal-curve, and implementation analyses.
- Data acquisition: Five in-house pediatric MRI acquisitions used a golden-angle stack-of-stars 3D FLASH sequence with 32 coronal slices and approximately 4.2–4.8-second temporal resolution.The protocol used 1.25 × 1.25 × 3 mm3 voxels and retrospectively binned 34 spokes per frame.
- Evaluation: Methods were qualitatively compared on representative aorta and kidney slices in both spatial and spatiotemporal views.Because no ground truth was available, PSNR and SSIM were computed against GRASP as reconstruction-similarity indicators.
- Dynamic analysis: Aorta and kidney signal curves were extracted from whole-volume reconstructions using segmentation masks while retaining the top 20% of mask intensities.Temporal behavior was additionally assessed using peak-to-baseline ratio (P/B), wash-in slope, and temporal variation after contrast arrival.
- Implementation: Optimization used adaptive density control, tier-specific temporal ranks, Adam, and 5000 iterations with a cosine learning-rate schedule.The base, contrast, and geometry tiers used temporal ranks Rb=3, Rc=4, and Rg=1, respectively.
4 Results
Gabor primitives produced sharper spatial and temporal reconstructions than the Gaussian variant and preserved temporal detail lost by Hash-INR. Gabor and Gaussian also retained the expected aortic enhancement peak and kidney-curve behavior, while the representation separated base, DCE, and motion dynamics.
- Reconstruction quality: Gabor recovered sharper spatial detail than Gaussian and preserved temporal sharpness that Hash-INR lost, while Hash-INR was slightly noisier and temporally blurred than GRASP.Quantitative image similarity supported the sharper reconstruction, although the passage cautions that these values are indicative only.
- Representation analysis: Gabor’s base tier stayed flat, its DCE tier rose at injection, and its motion tier learned general drift or respiratory-motion temporal bases.The analysis visualized the locations and scales of the top-20 enhanced primitives and temporal behavior of the top-10% enhancing primitives.
- Time-intensity curves: Around 35 s after injection, Gaussian and Gabor retained an aortic peak similar to GRASP, whereas Hash-INR and L+S oversmoothed it; kidney curves showed the same behavior.Table 1 reported similar peak behavior and post-contrast temporal smoothness for GRASP, Gabor, and Gaussian.
5 Discussion and Conclusion
The paper presents the first primitive-based framework for DCE-MRI reconstruction, using disentangled base, dynamic-contrast, and geometry-coupled tiers to reconstruct highly undersampled acquisitions competitively with GRASP while preserving sharp enhancement. Evaluation remains limited by the absence of ground truth, a small in-vivo cohort, and reconstruction costs that scale with temporal resolution.
- 5 Discussion and Conclusion: The first primitive-based DCE-MRI framework disentangles base, dynamic-contrast, and geometry-coupled tiers, achieving quality competitive with GRASP while preserving sharp aortic and renal enhancement.Recovered contrast primitives localize to expected anatomical regions.
- 5 Discussion and Conclusion: Evaluation is limited because GRASP temporally smooths the bolus, no ground-truth reference is available, and the study uses a small in-vivo cohort.Rendering every frame also couples reconstruction cost to temporal resolution, trading training time for finer dynamics and requiring further hyperparameter tuning.
- 5 Discussion and Conclusion: The explicit geometric representation supports higher acceleration through explicit regularization and motivates extensions to multidimensional MRI and tracer-kinetic parameter estimation.These downstream and broader MRI applications are identified as directions for further exploration.