Mathematical Methods in MRI - SFB MR-DYNAMO Workshop

20 May 2026, 07:05 → 22 May 2026, 12:45 Europe/Vienna

A-8010 Graz

1^st Workshop of the SFB MR-DYNAMO (FWF)

This workshop brings together applied mathematicians and MRI researchers working on the theoretical, computational, and algorithmic foundations of magnetic resonance imaging. In the spirit of the Mathematics of Reconstruction in Dynamical and Active Models (MR-DYNAMO) SFB, the workshop focuses on research that connects mathematical modeling, acquisition design, and image reconstruction across the full MRI pipeline. Further information on the SFB MR-DYNAMO is available at https://imsc.uni-graz.at/mr-dynamo/.

Topics of interest include inverse problems and optimal control in MRI, model-based and learning-based reconstruction, optimal experiment and acquisition design, motion-robust and dynamic imaging, physics-informed machine learning, uncertainty quantification, and the joint optimization of acquisition and reconstruction. The workshop aims to strengthen exchange between communities that often meet separately and to stimulate new collaborations at the interface of applied mathematics and MRI methodology.

Program Overview:

The workshop will be held in person at TU Graz on 20-22 May, 2026. The scientific program starts on Wednesday at 9:00 and ends on Friday at 12:45. A workshop dinner is planned for Thursday evening. Further details will be announced closer to the workshop.

Deadlines:

Registration: 30.04.2026
Abstract submission: 30.04.2026

Registration and Abstract Submission:

Attendance of the workshop is free of charge!
If you would like to participate in the workshop, please complete the Workshop Registration form. If you would also like to give a presentation, please complete the Abstract Submission form.

Poster:

The poster for the workshop is available in the Materials section below. 

Local Organizing Committee:

Silvia Lebosi
silvia.lebosi@uni-graz.at

Teresa Rauscher
teresa.rauscher@uni-graz.at

Richard Huber
richard.huber@uni-graz.at

Felix Glang
glang@tugraz.at

 

The Mathematical Methods in MRI - SFB MR-DYNAMO Workshop 2026 is supported by:

booklet.pdf

Datenschutzerklaerung_SFB_Workshop.pdf

workshop_announcement_poster.pdf

Wednesday 20 May

08:15 → 09:00 Registration

09:00 → 09:30 Opening

09:30 → 10:30 Plenary Talk: Prof. Moritz Zaiss

Convener: Prof. Martin Uecker (TU Graz)

09:30 MR-zero: How MR physicists move fast and break things

In the MR-zero project [1,2], we approached to reinvent MRI from scratch by treating MR sequence design as a learning-based optimization problem. We formulate an end-to-end differentiable simulation pipeline, given a specific imaging task or even just a target image, the entire chain can be optimized jointly.

In this talk, I will introduce the MR-zero approach and illustrate it with examples: learning MRI from scratch, improving image sharpness in TSE and FLASH MRI, optimizing sequences for imperfect scanner systems, and performing end-to-end quantitative MRI experiments. These results show what becomes possible when we borrow tools from optimization, automatic differentiation, and machine learning and combine it with recent MR physics developments like phase distribution graphs [3].

At the same time, MR-zero is far from an ideal mathematical framework. Many parts of the pipeline and problem formulation are still in progress, and are slow, unstable, approximate, or non-optimal. For this audience, this is exactly the point: the project exposes open problems in inverse problems in MR, differentiable simulation, typical optimization constraints and regularizations, and the translation from continuous models to executable scanner experiments. The talk is both a report and an invitation. I will show what we achieved with out-of-the-box mathematical tools, and where we would benefit from deeper mathematical insight. If you leave thinking, “What are they even doing? This could be done so much better,” then the talk has succeeded.

[1] https://mrsources.github.io/MRzero-Core/
[2] https://doi.org/10.1002/mrm.28727
[3] https://doi.org/10.1002/mrm.30055

Speaker: Prof. Moritz Zaiss (Friedrich-Alexander University Erlangen–Nuremberg (FAU))

10:30 → 11:00 Coffee Break

11:00 → 12:40 Contributed Talks: Session 1

Convener: Dr Felix Glang (Institute of Biomedical Imaging)

11:00 Direct signal control without Fourier coefficients

The well known direct signal control (DSC) approach allows us to optimize excitation voltages of a MRI pulse sequence in such a way that the overall transverse magnetization of a spin ensemble has desired values at a number of signal readout times. In this approach the overall transverse magnetization is calculated with the help of an approximate solution of the Bloch equations and corresponding transformations of a finite number of Fourier coefficients through the pulse sequence. We develop an equivalent alternative calculation method without Fourier coefficients. This formulation uses a finite number of discrete spins and the approximate solution of the Bloch eqquations for these spins. We show that the relevant Fourier coefficients of DSC can be obtained from a discrete Fourier transfom of the magnetizations of these spins. We specify how to select the positions and initial magnetizations of the discrete spins to get identical overall magnetizatios at the readout times as with DSC.

Speaker: Dr Kurt Majewski (Siemens AG, DAI R ORD-DE)

11:25 Arterial spin labeling MRI with radial sampling: end-to-end open-source sequence design and image reconstruction using BART

Introduction Arterial spin labeling (ASL) is a promising non-invasive perfusion imaging technique, but its reproducibility across platforms remains limited due to the lack of standardized, end-to-end pipelines. Although consensus guidelines exist, implementation variability persists. The Berkeley Advanced Reconstruction Toolbox (BART) is a free and open-source computational MRI framework for advanced image reconstruction that was recently extended to include pulse sequence design. We present a fully reproducible radial ASL pipeline in BART that integrates sequence generation, acquisition, and reconstruction. Radial sampling enables both conventional single-delay perfusion-weighted imaging (PWI) and continuous multi-delay acquisition within a single scan, improving time efficiency, motion robustness, and undersampling flexibility. To address ASL’s low signal-to-noise ratio, advanced reconstruction methods are incorporated. Methods A pseudo-continuous ASL (PCASL) sequence was implemented in BART following consensus recommendations. Data acquisition used a 2D radial sampling scheme on a 3T system in healthy volunteers, supporting both single-delay and continuous acquisitions. Image reconstruction employed parallel imaging compressed sensing with ASL-specific total generalized variation (ASL-TGV) regularization. Results and Discussion Single-delay acquisitions across multiple post-labeling delays (PLDs) demonstrated robust PWI generation, with ASL-TGV reducing noise at longer PLDs. Continuous acquisition provided improved structural detail and lower noise, especially at later time points, while capturing temporal dynamics efficiently. Notably, it achieved comparable temporal coverage in one quarter of the scan time and enabled retrospective temporal analysis. Conclusion This open-source radial ASL pipeline enables reproducible, vendor-independent workflows and flexible sequence sharing. Its modular design supports future extension to 3D imaging, promoting standardization and collaboration in ASL research.

Speaker: Ms Viktoria Buchegger (Institute of Biomedical Imaging, Graz University of Technology, Graz, Austria)

11:50 Open-source quantitative MRI: combined acquisition and reconstruction

This work presents an open-source framework that integrates MRI sequence design, execution, and model-based reconstruction into a single reproducible workflow within the BART toolbox. The motivation is to address reproducibility challenges in quantitative MRI (qMRI), where both sequence implementation and reconstruction details are often difficult to replicate due to proprietary software and incomplete methodological descriptions. The framework represents MRI sequences on the lowest level as events such as gradients, RF pulses, ADC readouts, and triggers. These events are combined into sequence blocks which represent a modular functionality, such as acquisition of a radial spoke. Blocks are a function of a controlable high-level parameter set. We propose two methods of acquisition. First, the sequence is generated prospectively offline, saved to a well-established open-source format (Pulseq) and executed via an appropriate interpreter program. Second, we integrate BART as a library dynamically on the scanner which allows for adjustment of parameters online while planning the scan and generates interpretable sequence events just-in-time during execution. In both cases, all necessary details of the acquisition, including timing, and k-space trajectory can be regenerated to ensure consistent reconstruction. We validated the framework with phantom and in-vivo experiments using radial sampling and model-based reconstruction in two advanced qMRI applications. First, subspace-based T1 mapping is performed using an inversion-recovery radial FLASH sequence with highly accelerated acquisition. Second, a multi-echo radial FLASH sequence is used for joint estimation of R2* and B0 maps. In both cases, parameter maps from different acquisition strategies show strong agreement, with small differences mainly attributed to noise or physiological variations. In summary, an end-to-end open-source framework for joint sequence design and model-based reconstruction in BART for full replicability of qMRI techniques is introduced. Both, a direct integration with online parameter adjustment on the scanner and an offline Pulseq export, is presented.

Speaker: Mr Daniel Mackner (Graz University of Technology)

12:15 MRI sequences: practical aspects of system integration

Open-Source software such as pypulseq and BART have immensely improved MRI sequence design and image reconstruction. Nevertheless, working with a clinical MRI scanner still poses challenges. We present solutions for bringing open source sequences and image reconstruction to the scanner. An MRI sequence describes the sequence of events that is played out by the scanner hardware during an MRI exam. This is generated by a software named the "sequence program". It implements a function mapping adjustable input parameters and hardware constraints to output events. The sequence program can be run ahead of scanning, but adaptation to specific patients/volunteers often requires running the sequence program on the scanner. Applications such as interactive MRI additionally require just-in-time generation of sequence events. Images can then be reconstructed from the measured data, by means of a reconstruction program. It can be run offline in a research setting, but a rapid and thus online reconstruction is often beneficial/mandatory. We have created a sequence program in BART that can run on the scanner by compiling BART to a shared library which exposes several functions, mainly: (de-)initialization, parameter updates, and calculation of sequence events. Another program, referred to as interpreter, translates BART and vendor API calls. Interactive updates are furthermore enabled using a non-blocking ring-buffer, keeping the just-in-time property. Real-time image reconstruction on the scanner is realized based on a protocol for streaming multidimensional arrays, which is complemented by an 'adapter' program that harnesses the vendor API, extracting and inserting raw-data / reconstructed images.

Speaker: Mr Philip Schaten (TU Graz)

12:40 → 14:05 Lunch Break

14:05 → 15:45 Contributed Talks: Session 2

Convener: Prof. Kristian Bredies (University of Graz)

14:05 On uniqueness of coefficient identification in the Bloch-Torrey equation for magnetic resonance imaging

In this talk we provide some uniqueness results for the (multi-)coefficient identification problem of reconstructing the spatially varying spin density as well as the spin-lattice and spin-spin relaxation times and the local field inhomogeneity in the Bloch-Torrey equation, as relevant in magnetic resonance imaging MRI.
To this end, we follow two approaches:
(a) Relying on sampling of the k-space and (approximately) explicit re-construction formulas in the simplified (Bloch) ODE setting, along with perturbation estimates;
(b) Relying on infinite speed of propagation due to diffusion.
The results on well-posendess and Lipschitz continuous differentiability of the coefficient-to-state map derived for this purpose, are expected to be useful also in the convergence analysis of reconstruction schemes as well in mathematical optimization of the experimental design in MRI.

Speaker: Prof. Barbara Kaltenbacher (University of Klagenfurt)

14:30 Operator splitting and adjoint-based L-BFGS for parameter identification in the Bloch-Torrey equation

We present a computational framework for model-based quantitative imaging through the identification of spatially varying relaxation parameters in the Bloch-Torrey equation, a time dependent PDE governing magnetization dynamics in magnetic resonance imaging. The equation couples RF induced precession, longitudinal and transverse relaxation, and diffusion advection transport, a structure that poses particular challenges for both discretization and adjoint derivation due to the interplay of rotational, dissipative, and elliptic operators acting on different time scales. The forward problem is discretized using a first order Lie operator splitting that preserves the physical structure of each sub process: precession is handled using the exact Rodrigues formula, relaxation using pointwise exponentials, and diffusion advection using an implicit P1 finite element scheme. The order of accuracy is rigorously verified using particular known solutions, confirming the expected temporal and spatial convergence rates. A key contribution is the derivation of a discrete adjoint consistent with the splitting scheme. In particular, the adjoint of the Rodrigues sub-step is its exact inverse rotation, a nontrivial requirement that ensures gradient consistency without resorting to continuous adjoints or automatic differentiation. Gradients of the least-squares cost functional are used within an L-BFGS method formulated in the L²(Ω) inner product, enabling distributed parameter identification for three spatially varying fields: the longitudinal rate R₁, the transverse rate R₂, and the equilibrium magnetization M_eq. Numerical experiments on a synthetic phantom demonstrate accurate simultaneous reconstruction of all three parameters from CPMG multi echo measurements, across a hierarchy of mesh refinements with warm-started interpolation between levels.

Speaker: Mr Pablo Muñoz (Klagenfurt Universität)

14:55 Observability as a constraint qualification for a nonsmooth bilevel PDE constrained problem

This talk gives a nontechnical introduction to PDE-constrained bilevel optimization problems with nonsmooth lower levels. The PDE constraints considered here are simplifications of the Bloch--Torrey equation from MRI, with the bilevel structure modelling a parameter learning problem. We discuss how complementarity conditions arising from first-order information are connected to the observability of an associated multiplier system. Based on the obtained multipliers, we propose a simple gradient-based method for computing the solutions of the bilevel problem.

Speaker: Dr Jyrki Jauhiainen (University of Graz)

15:20 Optimal control based estimation of blood pressure from MRI velocity data

Modern 4D-MRI imaging techniques are able to capture fully space-time resolved blood flow velocity data. This data facilitates a diagnosis of the severity of cardiovascular diseases, provided that, in a first step, additional information such as pressure, wall-shear stresses, etc., is inferred from the measured data. In this talk, we treat the inverse problem of deducing pressure values from MRI velocity measurements by an optimal control formulation. The flow inside a blood vessel is modeled by the transient Navier-Stokes equations. The equations are controlled by inhomogeneous Do-Nothing conditions imposed on the in- and outflow boundaries, which result from a truncation of the computational domain. While these conditions are widely used in numerics and are stable for pure outflow, instabilities arise when backflow occurs. On the continuous level, no global existence theory is available. Despite these difficulties associated with the Do-Nothing conditions, we present a formulation of a continuous optimal control problem, which is well posed. Our results thus bridge a gap in the existing literature, which so far has focused on numerical implementations of related optimal control problems. We show numerical examples that highlight the potential of our approach.

Speaker: Mr Jakob Wagner (Technical University of Munich)

15:45 → 16:15 Coffee Break

16:15 → 17:30 Contributed Talks: Session 3

Convener: Prof. Martin Uecker (TU Graz)

16:15 Nonlinear gradient modulations, multi-shot EPI, parallel imaging in a unified RKHS framework

MRI mainly relies on linear gradients for spatial encoding. However, nonlinear spatial encoding functions – such as RF receivers’ sensitivities in parallel imaging, actively produced nonlinear gradient modulations in FRONSAC, or naturally occurred shot-to-shot phase variations in multi-shot EPI – can significantly impact acquisition speed and spatial resolution. Inspired by the reproducing kernel Hilbert space (RKHS) framework used to quantify signal sampling of parallel imaging directly in k-space, we proposed to manage diverse nonlinear spatial encoding functions in a unified mathematical perspective based on RKHS. Under such view, we presented several techniques across two distinct areas to achieve robust, fast, and high-resolution MRI. In one example, we generate nonlinear gradient modulations using an 8-channel local B0 coil array in a 9.4T human scanner. Applying sinusoidal modulations to 8 independent B0 channels during FLASH readout to accelerate sampling, we can now encode, auto-calibrate, compress, and reconstruct such sampled data entirely in k-space – analogous to handling multi-coil data in parallel imaging. Another example is to auto-calibrate the shot-to-shot instability in multi-shot EPI. Specifically, we introduce a small trajectory overlap between shots. By applying a GRAPPA/ESPIRiT type operation to these overlap regions, shot-to-shot phase variation kernels/maps can be extracted, enabling robust self-navigation for various multi-shot trajectories. Ultimately, under this RKHS view, nonlinear gradient modulations, multi-shot EPI, and parallel imaging can be mathematically bridged much more thoroughly. Unifying these physically distinct areas allows them to share the well-established efficiency and robustness of parallel imaging, offering new possibilities for fast and high-resolution MRI.

Speaker: Mr Rui Tian

16:40 Non-standard inner products in MRI reconstruction

This abstract aims to describe MRI reconstructions in the general framework of finite-dimensional inner-product spaces. We explain that image-space preconditioning (ISP) and data-space preconditioning (DSP) can be formulated as non-conventional inner-products. This allows to introduce image-space preconditioning in the variational formulation of the MRI reconstruction problem (in an algorithm-independent way) and to propagate it in principle in all iterative reconstructions, including many iterative deep-learning and compressed-sensing reconstructions where preconditioning is lacking until now.

Speaker: Dr Bastien Milani

17:05 Physical correctness of data consistency in state space model-based MRI reconstruction

State space models (SSMs) have recently emerged as efficient alternatives to Transformers for MRI reconstruction, offering linear complexity while maintaining global receptive fields. LMO (Linear Mamba Operator, Li et al., CVPR 2025) formulates MRI reconstruction as operator learning in bandlimited function spaces, achieving competitive results on standard benchmarks. We identify a physical modeling error in LMO's data consistency (DC) mechanism: the DC layer operates on magnitude images, discarding phase information before enforcing k-space consistency with complex-valued measured data. This mixes incompatible frequency-domain representations at every DC step, introducing systematic errors throughout the network. Notably, this behavior is undocumented — it is only visible in the released code. We propose a physically grounded correction that preserves the SSM architecture: (1) sensitivity-weighted complex input (2-channel real/imaginary), (2) a ComplexDC layer applying proper complex-valued k-space enforcement, and (3) adapted lifting and projection layers — adding only 1.4% parameters. On FastMRI Brain (Poisson 2D, 8x acceleration), the corrected model yields +1.74 dB PSNR, +0.012 SSIM, and -24% NMSE over the magnitude-domain baseline, with only 2.15M parameters — less than half of the original LMO (~5M) and a fraction of UNet (31M). Our results suggest that physical correctness of data consistency is a more effective lever for reconstruction quality than architectural refinements to the SSM itself, particularly in data-limited regimes.

Speaker: Mr Xin Zhao (Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences)

Thursday 21 May

09:00 → 10:00 Plenary Talk: Prof. Michael Unser

Convener: Prof. Thomas Pock (TU Graz)

09:00 Deep-spline neural networks for stable image reconstruction

The two dominant CNN-based paradigms for consistent biomedical image reconstruction are: (i) approaches based on trainable regularizers; and (ii) proximal-gradient-type architectures, such as plug-and-play (PnP) methods, that rely on a learned denoiser. A major challenge in both settings is the control of stability and convergence, which typically requires restrictive conditions such as convexity of the regularizer or nonexpansiveness of the denoiser. While effective from a theoretical standpoint, these constraints may reduce the representational power of the resulting models.

In this work, we propose a framework that increases model flexibility by learning the neuronal activation functions jointly with the network parameters, while maintaining explicit control over their slopes. The activations are parameterized as adaptive linear splines obtained through a second-order total-variation regularization principle. This formulation leads to parsimonious and trainable spline-based neural architectures that can be integrated within stable reconstruction schemes. We demonstrate the practicality of the proposed approach on denoising and biomedical image-reconstruction tasks, where it achieves competitive and robust performance.

Speaker: Prof. Michael Unser (Ecole Polytechnique Fédérale de Lausanne (EPFL))

10:00 → 10:50 Contributed Talks: Session 4

Convener: Prof. Thomas Pock (TU Graz)

10:00 Dynamic transitions for fast joint acquisition and reconstruction of CEST-Rex and T1

This work proposes a method for the simultaneous estimation of the exchange-dependent relaxation rate Rex and the longitudinal relaxation time T1 from a single acquisition. A novel acquisition scheme was developed that combines CEST saturation with an inversion pulse and a Look-Locker readout to capture the magnetization evolution starting from the inverse transient Z-spectrum. The corresponding signal model, derived from the Bloch-McConnell equations, describes both the transient Z-spectrum and the Look-Locker dynamics. A model-based reconstruction approach is employed to jointly estimate Rex and T1. The proposed method was validated using a numerical phantom and benchmarked against conventional CEST and Look-Locker T1 mapping in a phantom and in vivo on a clinical 3T scanner. The joint estimation approach demonstrated strong agreement with ground truth and conventional methods across a wide range of T1 and CEST parameters. The acquisition time was reduced by 20%–30% compared to standard CEST protocols, while providing a higher signal-to-noise ratio (SNR) in parameter maps. In conclusion the proposed technique enables robust and efficient simultaneous quantification of CEST Rex and T1in a single acquisition. It improves parameter map quality and reduces scan time, making it suitable for both phantom and in vivo imaging across a wide range of physiological conditions.

Speaker: Dr Markus Huemer (Institute of Biomedical Imaging - Graz University of Technology)

10:25 Learning multi-pool dynamics in magnetic resonance imaging

Magnetic Resonance Imaging (MRI) is a key non-invasive imaging modality offering large versatility in creating high-resolution images. Reconstructing MR images from measurements requires knowledge of the MR physics involved in the measurement process, which are highly complex (e.g., involving quantum-mechanical effects) and are commonly modeled via the Bloch equations. Recent investigations have shown that the consideration of multi-pool models using Bloch-McConnell equations are beneficial, however, the related physics are still not fully understood. We address the challenge of uncovering hidden physics in MRI through a structured model learning approach. Specifically, we build upon the fundamental Bloch model and propose its extension with a potentially non-linear source term that captures the unknown dynamics associated with the multi-pool structure of the Bloch-McConnell model. The source term is formulated as a recurrent neural network with a highly interpretable architecture, strongly inspired by principles derived from the Bloch-McConnell model. We demonstrate the versatility and generalizability of the proposed framework and evaluate the learning process using artificially generated data in a multi-pool setting.

Speaker: Mr Štěpán Zapadlo (University of Graz)

10:50 → 11:20 Coffee Break

11:20 → 12:35 Contributed Talks: Session 5

Convener: Prof. Christian Clason (University of Graz)

11:20 An inertial Langevin algorithm

We present a novel method for drawing samples from Gibbs distributions with densities of the form π(x)∝exp(−U(x)). The method accelerates the unadjusted Langevin algorithm by introducing an inertia term similar to Polyak's heavy ball method, together with a corresponding noise rescaling. Interpreting the scheme as a discretization of kinetic Langevin dynamics, we prove ergodicity (in continuous and discrete time) for twice continuously differentiable, strongly convex, and L-smooth potentials and bound the bias of the discretization to the target in Wasserstein-2 distance. In particular, the presented proofs allow for smaller friction parameters in the kinetic Langevin diffusion compared to existing literature. Moreover, we show the close ties of the proposed method to the over-relaxed Gibbs sampler. The scheme is tested in an extensive set of numerical experiments covering simple toy examples, total variation image denoising, and the complex task of maximum likelihood learning of an energy-based model for molecular structure generation. The experimental results confirm the acceleration provided by the proposed scheme even beyond the strongly convex and L-smooth setting.

Speaker: Mr Alexander Falk (TU Graz)

11:45 Preconditioned Langevin sampling for Bayesian imaging

In this talk we consider time- and space-dependent preconditioning for Langevin sampling from posteriors arising in Bayesian inverse imaging problems, in particular, MRI. Inspired by quasi-Newton methods in optimization, we consider learned preconditioners which use curvature information to accelerate sampling. We provide a careful theoretical analysis of the sampling methods based on the discretization of the corresponding time-inhomogeneous preconditioned Langevin diffusion as well as experimental results for small scale problems and Bayesian imaging.

Speaker: Dr Andreas Habring

12:10 Fast and robust diffusion posterior sampling for MR image reconstruction using the preconditioned unadjusted Langevin algorithm

Introduction Diffusion models employed as prior knowledge have demonstrated strong performance in MRI reconstruction. A principal advantage of this probabilistic framework is that the variability across generated samples enables uncertainty quantification. Incorporating the likelihood term directly into the diffusion process, however, yields intractable expressions that are typically approximated using various methods. In this work, we examine the explicit inclusion of the likelihood term and propose the use of the exact likelihood, unaltered by the diffusion process. In addition, we include preconditioning into the Unadjusted Langevin Algorithm (ULA) to achieve fast convergence without step size tuning over different MRI reconstruction problems. Methods A UNet trained on the fastMRI brain dataset via score matching is used as a prior. The likelihood is combined with this prior at each noise level, and sampling is performed using the preconditioned ULA (pULA). Inference is conducted on both Cartesian and radial brain data from a healthy volunteer, acquired on an in-house Siemens 3T scanner, with retrospective undersampling at acceleration factors of 4 and 8. We compare pULA against the standard annealed likelihood approach proposed by Jalal et al. and diffusion posterior sampling introduced by Chung et al. Results pULA outperforms both methods in terms of PSNR and SSIM. In addition the proposed method is more robust to the choice of step size and converges faster than the other methods. Discussion and Outlook The proposed method shows fast and robust posterior sampling over various reconstruction problems outperforming current state-of-the-art methods.

Speaker: Ms Tina Holliber (Graz University of Technology, Institute of Biomedical Imaging)

12:35 → 13:30 Lunch Break

13:30 → 14:30 Hands-on MRI

We will give a live demonstration of real-time cardiac MRI at the 3T MRI scanner of TU Graz, as well as imaging using a self-built portable low-field MRI scanner.

14:30 → 15:20 Contributed Talks: Session 6

Convener: Prof. Martin Holler (University of Graz)

14:30 Zero-shot self-supervised learning of spatio-temporally varying regularisation parameter maps for dynamic cardiac MR image reconstruction

Data-driven reconstruction methods, mostly based on deep learning, have emerged as the undisputed state-of-the-art across many different inverse problems, including MRI reconstruction. However, they are also notorious for requiring large and diverse datasets to be trained successfully, preferably in a supervised manner. When target reconstructions cannot be obtained, zero-shot self-supervised learning approaches have been proposed as adequate alternatives for model training, often requiring some additional mechanisms to avoid overfitting. In this work, we consider a zero-shot self-supervised training approach for a hybrid reconstruction method that is based in the combination of a hand-crafted regulariser and a deep neural network that adapts, in a spatio-temporal manner, the strength of the employed prior. We demonstrate that this combination allows for fast learning of the regularisation strength, achieving nearly the same performance as supervised pre-training combined with sample-wise, test-time self-supervised training. Further, we show that the combination of a hand-crafted prior with a learned adaptive strength systematically avoids overfitting, eliminating the need for strategies such as monitoring metrics on validation sets or early stopping. Finally, the considered methods are highly interpretable due to the choice of the hand-crafted priors. We showcase the proposed paradigm for spatio-temporally adaptive Total Variation (TV) and Total Generalised Variation (TGV) applied to image denoising and dynamic cardiac MRI reconstruction.

Speaker: Dr Kostas Papafitsoros (Queen Mary University of London)

14:55 Gabor primitives for accelerated cardiac cine MRI reconstruction

Accelerated cardiac cine MRI requires reconstructing spatiotemporal images from highly undersampled k-space data. Implicit neural representations (INRs) enable scan-specific reconstruction without large training datasets, but encode content implicitly in network weights without physically interpretable parameters. Gaussian primitives provide an explicit and geometrically interpretable alternative, but their spectra are confined near the k-space origin, limiting high-frequency representation. We propose Gabor primitives for MRI reconstruction, modulating each Gaussian envelope with a complex exponential to place its spectral support at an arbitrary k-space location, enabling efficient representation of both smooth structures and sharp boundaries. To exploit spatiotemporal redundancy in cardiac cine, we decompose per-primitive temporal variation into a low-rank geometry basis capturing cardiac motion and a signal-intensity basis modeling contrast changes. Experiments on cardiac cine data with Cartesian and radial trajectories show that Gabor primitives consistently outperform compressed sensing, Gaussian primitives, and hash-grid INR baselines, while providing a compact, continuous-resolution representation with physically meaningful parameters.

Speaker: Mr Wenqi Huang (Technical University of Munich)

15:20 → 15:50 Coffee Break

15:50 → 16:40 Contributed Talks: Session 7

Convener: Dr Richard Huber (University of Graz)

15:50 Geometrically-grounded 3D+time representations from sparse 2D MR views

Recovering high-fidelity 3D volumes from sparse or degraded 2D images is a fundamental challenge in magnetic resonance imaging with broad clinical applications. Implicit neural representations (INRs) offer a resolution-agnostic solution to modelling a volume directly from a scanner’s coordinate system. This allows for continuous 3D representations to be built from arbitrarily oriented sets of imaging slices, as the exact position and orientation of every voxel can be derived from DICOM headers. Modelling volumes directly on the spatial geometry allows us to impose physical constraints into the learning process. We can employ the point-spread function to accurately model the acquisition physics of a given anisotropic voxel. Moreover, inter-slice subject motion can be corrected in an unsupervised manner by expressing a slice’s rigid transformation as optimizable translation and rotation parameters. These methods are also capable of handling multi-subject cohorts, allowing for statistical correlations on tissue position to be exploited across a population. In our recent works, we apply these methods to a multi-subject cohort of 2D+time cardiac CINE datasets. We demonstrate that, by training on intensity-segmentation pairs, these representations can interpolate 3D+time volumes at any desired resolution. At test-time deriving a new subject’s representation from imaging information alone provides free complimentary segmentation labels. Ongoing work includes applying further physical constraints. The time dimension can be framed under a velocity-based deformation process, enforcing physically-grounded tissue motion. Furthermore, exchanging the neural representation in favor of Gaussian primitives allow us to formulate analytically-computable versions of the aforementioned constraints, drastically reducing compute costs.

Speaker: Mr Nil Stolt-Ansó (Technical University of Munich)

16:15 Patient-specific 3D ventricular models from cardiac cine MRI: a segmentation, registration, and volumetric modeling framework

Accurate 3D cardiac model construction relies on high-quality segmentation and registration of 2D+t cardiac cine MRI (cMRI) data, with patient-specific ventricular models serving as a key building block for cardiac digital twins. 2D+t cMRI offers higher resolution compared to 3D acquisitions, but the 2D slices are prone to spatial misalignment caused by patient movement and inconsistent breath holds and heart beats during acquisition. Left uncorrected, such misalignments propagate into the 3D modeling stage, degrading anatomical accuracy. We present a semi-automated, modular pipeline for segmenting and registering 2D+t cMRIs and constructing personalized 3D anatomical models of the left and right ventricles and left ventricular myocardium (LV, RV, MYO). Data from the 'ILearnHeart' project consist of 2D+t cMRI scans in 2-chamber (2CH), 4-chamber (4CH), and stacked short-axis (SAX) views from seven healthy subjects, covering a full cardiac cycle. We segment the LV and RV blood pools and LV myocardium using nnU-Net and a Gaussian process-based few-shot approach. Registration proceeds in two steps: intensity-based alignment of 2CH and 4CH slices via normalized cross-correlation (NCC), followed by contours-based alignment of the SAX stack through intersection distance minimization. Intensity-based registration yields an average NCC improvement of (17.5 ± 16.7)% over unregistered slices. The registered contours serve as input for 3D surface reconstruction via NURBS lofting, yielding smooth, patient-specific ventricular surface models of the LV, RV, and MYO. Future work targets closed volumetric mesh generation and 4D extension across the cardiac cycle.

Speaker: Dr Kathrin Lisa Kapper (University of Graz, Austria)

18:30 → 21:30 Conference Dinner

Friday 22 May

09:00 → 10:00 Plenary Talk: Prof. Gerlind Plonka-Hoch

Convener: Prof. Barbara Kaltenbacher (University of Klagenfurt)

09:00 The multichannel blind deconvolution problem in parallel MRI

One of the biggest innovations in magnetic resonance imaging (MRI) within the last years was the concept of parallel MRI. In this setting, the use of multiple receiver coils allows the reconstruction of high-resolution images from undersampled Fourier data such that the acquisition time can be substantially reduced. Mathematically, the parallel MRI reconstruction problem can be seen as a multi-channel blind deconvolution problem, where the coil sensitivity functions and the magnetization image have to be recovered simultaneously from the acquired data.

In this talk, we will give a short survey on the mathematical background of some existing reconstruction methods in parallel MRI so far, including GRAPPA, SPIRiT, and ESPIRiT.

Further, we propose the MOCCA algorithm to recover the coil sensitivities and the magnetization image from incomplete Fourier measurements. Our approach is based on a parameter model for the coil sensitivities using bivariate trigonometric polynomials of small degree.

The derived MOCCA algorithm provides perfect reconstruction results if the model assumptions are satisfied. Moreover, it has a low computational complexity and fits real MRI data sufficiently well such that it is applicable in practice.

The results presented in this talk have been obtained jointly with Yannick Riebe and Benjamin Kocur.

Speaker: Prof. Gerlind Plonka-Hoch (University of Göttingen)

10:00 → 10:50 Contributed Talks: Session 8

Convener: Prof. Barbara Kaltenbacher (University of Klagenfurt)

10:00 MRI-NUFFT: an open-source toolbox for high-performance non-Cartesian reconstruction and trajectory design

MRI reconstruction from non-Cartesian k-space undersampling is naturally formulated as an inverse problem (y = Ax + n), where recovering the image x from data y critically depends on the accuracy and efficiency of the forward model A. MRI-NUFFT is an open-source Python library that provides a unified, high-performance implementation of this physics-based forward model for non-Cartesian MRI using the Non-Uniform Fast Fourier Transform (NUFFT). It additionally offers multi-coil sensitivity estimation (ESPIRIT), density compensation strategies, and off-resonance correction. Centered on the core operator A, MRI-NUFFT enables efficient computation of forward, adjoint, and pseudo-inverse mappings across multiple backends and hardware. It supports automatic differentiation with respect to both data and physics parameters (k-space trajectories, field maps, sensitivity maps), native interoperability with NumPy, CuPy, and PyTorch, and seamless CPU/GPU execution, making it well-suited for optimization-based reconstruction and learning-based frameworks. A distinctive feature is its tight integration between reconstruction and acquisition design: MRI-NUFFT includes a rich collection of parameterized 2D and 3D non-Cartesian trajectories, supports hardware-constrained trajectory projection, and enables gradient waveform design and optimization under magnetic gradient hardware constraints. Compatibility with Pulseq further facilitates the generation of executable sequences, bridging simulation and scanner deployment. By positioning the forward model as a central, reusable component, MRI-NUFFT enables rapid prototyping of inverse problem formulations, from variational methods to physics-informed machine learning and joint optimization of sampling and reconstruction.

Speaker: Mr Pierre-Antoine Comby (CEA/Neurospin)

10:25 Neural network-based motion-aware reconstruction of dynamic MRI data

Magnetic resonance imaging (MRI) has become a standard diagnostic tool in medical assessments. However, long acquisition times remain one of its main drawbacks. This makes dynamic imaging, as needed in cineMRI, particularly challenging: motion artefacts and limited data acquisition per time point lead to an inherently ill-posed reconstruction problem. Typical solution strategies include breath-holding by the patient to limit induced motion or manual binning of frames according to an external signal. We propose a method that incorporates motion signal information directly in the image discretisation. Inspired by the deep-image prior, invertible residual networks, Lipschitz-constrained networks via Householder reflections, and previous work on motion disentanglement, we design a multi-component neural network-based architecture for image reconstruction. The motion signals, in this case representative of the breathing and cardiac cycle, are used as conditioning information for the network. This framework not only allows for retrospective modification and suppression of motion but also enables the reconstruction of incomplete motion information, such as when no breathing belt is used. In our presentation, we provide a mathematically sound formulation of the reconstruction problem within a statistical learning framework, together with general features of the proposed network architecture. These are complemented by numerical test cases highlighting capabilities such as cine reconstruction, retrospective motion suppression, feature tracking, displacement extraction, and resolution independence.

Speaker: Mr Matthias Höfler (University of Graz)

10:50 → 11:20 Coffee Break

11:20 → 12:10 Contributed Talks: Session 9

Convener: Mr Markus Huemer (TU Graz)

11:20 On the role of smoothness constraints in MR fingerprinting sequence optimization

Magnetic Resonance Fingerprinting (MRF) allows to determine MR relaxation times quantitatively, but its performance depends on the design of acquisition schemes to enable robust parameter estimation from highly undersampled data [1]. Earlier work by Zhao et al. introduced a smoothness constraint in a Cramér-Rao lower bound-based optimization framework, motivated by the observation that smoother flip angle sequences show better performance with respect to undersampling artifacts [2]. To reduce undersampling-related artifacts, several sequence optimization approaches have been proposed [3,4]. Notably, these methods often impose an additional smoothness constraint on the flip angle sequence, even when the optimization objective already explicitly accounts for undersampling effects. In this work, we investigate the role of smoothness regularization in MRF sequence optimization. This raises a fundamental question: why is smoothness regularization still required in optimization schemes that are already designed to mitigate undersampling artifacts? To address this, we analyze the effect of different smoothness constraints on the resulting sequences and their performance in numerical simulations, phantom measurements, and in vivo experiments. Our results show that the commonly used smoothness constraint is not optimal in this setting and that alternative choices can improve performance. Furthermore, we demonstrate that omitting smoothness regularization altogether leads to a breakdown of the optimization procedure, underscoring its importance for stable and effective sequence design in MRF. Ultimately, this shows that the assumptions used to describe the functional to be minimized are overly simplistic and do not lead to improvements without additional constraints. 1 Ma D et al. Magnetic resonance fingerprinting. Nature. 2013;495:187–192. 2 Zhao B et al. Optimal experiment design for magnetic resonance fingerprinting: Cramér-Rao bound meets spin dynamics. IEEE Trans Med Imaging. 2019;38(3):844–861. 3 Heesterbeek DGJ et al. Mitigating undersampling errors in MR fingerprinting by sequence optimization. Magn Reson Med. 2023;89:2076–2087 4 Zibetti MVW et al. Optimization of 3D MR fingerprinting with Cramér-Rao lower bound and smooth signal evolutions for efficient T1, T2, and T1ρ in the knee joint. In: Proceedings of ISMRM Annual Meeting; 2025 May; Honolulu, Hawaii, USA. Abstract 1688.

Speaker: Mr Tim Höpfner (University Hospital Würzburg)

11:45 MRI2Qmap: multi-parametric quantitative mapping using multimodal contrast-weighted MRI denoising priors

Magnetic Resonance Fingerprinting (MRF) and other highly accelerated transient-state parameter mapping techniques enable simultaneous quantification of multiple tissue properties, but often suffer from aliasing artifacts due to compressed sampling. Incorporating spatial image priors can mitigate these artifacts, and deep learning has shown strong potential when large training datasets are available. However, extending this paradigm to MRF-type sequences remains challenging due to the scarcity of quantitative imaging data for training. Can this limitation be overcome by leveraging sources of training data from clinically-routine weighted MRI images? To this end, we introduce MRI2Qmap, a quantitative reconstruction framework that integrates the physical acquisition model with priors learned from deep denoising autoencoders pretrained on large multimodal weighted-MRI datasets. MRI2Qmap demonstrates that spatial-domain structural priors learned from independently acquired datasets of routine weighted-MRI images can be effectively used for quantitative MRI reconstruction. The proposed method is validated on highly accelerated 3D whole-brain MRF data from both in- vivo and simulated acquisitions, achieving competitive or superior performance relative to existing baselines without requiring ground-truth quantitative imaging data for training. By decoupling quantitative reconstruction from the need for ground-truth MRF training data, this framework points toward a scalable paradigm for quantitative MRI that can capitalize on the large and growing repositories of routine clinical MRI. (more details at: https://arxiv.org/abs/2603.11316)

Speaker: Dr Mohammad Golbabaee (University of Bristol)

12:10 → 12:20 Closing