Source-linked AI summary
Conditional Flow Matching for Cross-Field MRI Harmonisation
Baris Imre, Aram Salehi, Levente Baljer, Andrew Webb, Marius Staring, Efe Ilicak
TL;DR
Cross-field MRI differences complicate comparison and pooling, so the paper develops a single conditional flow-matching model for directed translation across five field strengths and three contrasts. It transports registered source slices directly to targets, using staged training from unpaired restoration data to paired adaptation and adversarial refinement. The model reaches mean SSIM 0.909 on the challenge evaluation set and outperforms matched regression and diffusion baselines on all three challenge metrics.
Problem
MRI images vary across field strengths and hardware, complicating comparison, reproducibility, and pooling across sites.
Method
A single conditional flow-matching model transports registered source slices directly to targets, with unpaired restoration pretraining, paired cross-field finetuning, and adversarial refinement.
Results
Mean SSIM 0.909 was achieved on the challenge evaluation set, outperforming regression and diffusion baselines on all three challenge metrics.
Takeaways & Limitations
One 6.3M-parameter model covers all 20 directed field pairs across three contrasts, using five solver steps per slice at inference.
Abstract
from arXiv · showhide
Magnetic resonance images of the same subject look markedly different across field strengths, which complicates the comparison and pooling of data across sites. We address cross-field brain-MRI translation for the MRIxFields2026 challenge, and in particular its Task~3: a single model that translates between any directed pair of the five field strengths and across three contrasts. We phrase the problem as a conditional flow matching path: because the source and target volumes are spatially registered, we learn a velocity field that carries the source slice directly to the target slice, rather than starting from noise. To learn this mapping from only three paired subjects, the unified model is trained in three stages: a degradation-bridge pretraining that distills a restoration prior from the abundant unpaired retrospective cohort, a cross-field finetuning over all directed pairs on the paired cohort, and an adversarial refinement that sharpens the output. At inference, we integrate the learned velocity with a second-order Heun solver in a handful of steps. A restoration prior learned without any paired data already reaches a mean SSIM of 0.837, and each subsequent training stage improves on it. A single 6.3M-parameter model thereby covers all 60 field-pair and contrast combinations, with inference in five solver steps per slice. On the challenge evaluation set the model reaches a mean SSIM of 0.909, averaged over the three contrasts, outperforming regression and diffusion baselines built on the identical network on all three challenge metrics.
1 Introduction
MRI appearance varies with field strength and acquisition setting, hindering comparison, reproducibility, and data pooling. The paper addresses this through a unified conditional-flow model for any directed field pair and three contrasts, trained with staged use of paired and unpaired data.
- MRI images are difficult to compare across sites and hardware, hindering reproducibility and pooling for large studies.
- Field strength changes signal-to-noise ratio, achievable resolution, and tissue contrast, making scans of the same subject look different across fields.
- The model targets Task 3 with one conditional flow-matching system translating between any directed pair of five field strengths across three contrasts.
- Training combines degradation-bridge pretraining on unpaired data, cross-field finetuning on paired data, and adversarial refinement.
2 Materials and Methods
The method learns deterministic source-to-target transport between registered MRI slices with conditional flow matching, then integrates the learned velocity using Heun’s method. A restoration prior from unpaired data is adapted through paired cross-field finetuning and adversarial refinement.
- Conditional flow matching: Conditional flow matching regresses a time-dependent velocity field that transports paired source samples to target samples along a deterministic path.The paper uses linear interpolation between paired endpoints, with velocity x1−x0 as the regression target.
- Conditional flow matching: Registered source and target slices are bridged directly instead of starting from noise, with the source provided as the integration’s t = 0 state.The network receives the interpolated state rather than the source as a separate input, keeping the bridge leak-free.
- Neural network conditioning: The network conditions a shared model on timestep, source field, and target field, combining embeddings into a vector that modulates features through FiLM.The timestep uses a sinusoidal embedding, while source and target fields use learned embedding tables.
- Inference: Inference integrates the learned velocity from t = 0 to t = 1 with a second-order Heun predictor–corrector scheme.Heun predicts with an Euler step and averages velocities at the current and predicted states, allowing fewer integration steps than a single Euler step.
- Adversarial refinement: Adversarial refinement targets realistic local texture by adding a conditional multi-scale PatchGAN discriminator to the flow-matching objective.A velocity-regression faithfulness anchor is retained alongside adversarial and feature-matching terms.
- Staged training: Three-stage training uses unpaired degradation-bridge pretraining, paired cross-field finetuning over ordered field pairs, and paired adversarial refinement.Pretraining learns a restoration prior; finetuning adapts it to directed mappings, while refinement rolls out without target conditioning and backpropagates through the trajectory.
3 Experiments and Results
Experiments evaluate a 6.3M-parameter slice-based model against matched regression and diffusion baselines on the challenge set. Ablations show gains from staged training and that five Heun steps recover the refined model’s best performance.
- 3.1 Implementation details: The 6.3M-parameter velocity network is a 2D U-Net operating on filtered, independently normalised axial slices.Three-dimensional volumes are treated as stacks of axial slices, with empty slices removed before scaling to the network range.
- 3.1 Implementation details: The regression and diffusion baselines share the U-Net, FiLM conditioning, data splits, and two-stage schedule, isolating the training objective.Regression uses a single forward pass, whereas diffusion denoises a noised source in 50 deterministic DDIM steps.
- 3.2 Results and ablation: 0.909 SSIM is achieved by the final model on the challenge evaluation set, averaged across 20 directed field pairs and three contrasts.The evaluation also reports LPIPS and nRMSE, with aggregate metrics across three held-out subjects per directed pair.
- 3.2 Results and ablation: 0.817 SSIM with one Heun step rises to 0.909 with five steps, while ten steps provide no further gain.The refined velocity is trained through a multi-step rollout, calibrating it for multi-step integration.
4 Discussion and Conclusion
The paper presents a single conditional-flow-matching model for registered cross-field brain-MRI translation across all directed field pairs and three contrasts. Its three-stage training recipe uses unpaired data for restoration pretraining, paired data for directed mapping, and adversarial refinement for sharper outputs.
- 4 Discussion and Conclusion: A single 6.3M model bridges registered source slices to targets across 20 directed field pairs and three contrasts.The model conditions its velocity field on the translation setting rather than using separate models for each pair.
- 4 Discussion and Conclusion: The training recipe combines degradation-bridge pretraining, cross-field finetuning, and adversarial refinement.The stages respectively use unpaired retrospective data, paired travelling-subject data, and refinement of the finetuned generator.
- 4 Discussion and Conclusion: The final model reaches mean SSIM 0.909, LPIPS 0.089, and nRMSE 0.227 over every directed pair and contrast.It outperforms architecture-matched direct regression and noised-source diffusion baselines.