MRIxFields2026: A Generalizable Cross- Field MRI Translation and Harmonization Challenge

Magnetic resonance imaging (MRI) plays a central role in neuroscience research and clinical assessment of brain structure and function. However, the rapidly expanding spectrum of MRI field strengths, from 0.1T to 7T scanners, has introduced substantial heterogeneity in signal-to-noise ratio, contrast behavior, spatial resolution, and B0/B1 field uniformity. These field-dependent differences fundamentally limit data comparability across centers and pose a major barrier to the generalization and clinical translation of AI-based MRI models.

Despite growing interest in cross-field MRI synthesis and enhancement, the field lacks a standardized, large-scale benchmark that spans the full range of magnetic field strengths and enables systematic evaluation of model robustness, structural fidelity, and field-awareness. Existing datasets and challenges are typically confined to narrow field ranges or single-domain settings, making it difficult to assess whether generative models can generalize across realistic multi-field and multi-center scenarios.

This challenge introduces the first comprehensive public benchmark for multi-field brain MRI image generation, covering field strengths from 0.1T to 7T and incorporating two major structural MRI contrasts (T1w,T2w, and T2-FLAIR). The dataset comprises approximately 500 retrospective cases and 200 prospectively acquired cross-field paired scans, with unified acquisition protocols, standardized preprocessing, and rigorous quality assurance. The paired multi-field acquisitions provide a unique reference for evaluating structural consistency across field transformations. 

Building on this dataset, the challenge defines three complementary and clinically motivated tasks:

1. Ultra-High Field MRI Synthesis from Arbitrary Magnetic Field Strengths The first task targets the generation of high-field-equivalent MRI from arbitrary input field strengths, enabling models to recover fine anatomical details and quantitative properties associated with 7T imaging. This capability is essential for research centers and early-stage clinical studies where ultrahigh-field scanners are scarce but high-quality representations are indispensable.
2. Higher-Field MRI Generation from Ultra-Low Magnetic Field Strengths The second task addresses a rapidly emerging global priority: enhancing ultra-low-field MRI to restore clinically meaningful tissue contrast under severely degraded imaging conditions. As 0.1T–0.3T systems become critical tools for low-resource settings, emergency medicine, and point-of-care imaging, AI-based enhancement becomes a necessary bridge to achieve diagnostic reliability without expensive hardware.
3. Controllable Field-to-Field MRI Synthesis with a Unified Conditional Model The third task introduces controllable, generalizable field-to-field synthesis via explicit conditioning mechanisms. This unified modeling strategy is crucial for harmonizing datasets across hospitals, reducing domain shift in multi-center studies, supporting federated learning environments, and enabling large-scale clinical AI deployment across heterogeneous scanners.

Evaluation combines voxel-level accuracy (nRMSE), structural similarity (SSIM), perceptual plausibility (LPIPS), and anatomically grounded segmentation metrics, including mean Dice overlap and normalized volume consistency across 14 bilateral deep gray matter nuclei derived from paired travelling-volunteer scans. These five complementary metrics jointly define the primary leaderboard via rank-summation, enabling balanced assessment of numerical fidelity, perceptual realism, and regional anatomical preservation. Top-ranked submissions will additionally undergo blinded neuroradiologist review to identify hallucinated anatomy or implausible structural alterations.

Together, these tasks establish a unified evaluation framework that moves beyond isolated domain translation and directly tests field-aware and controllable MRI generation. The challenge aims to advance both methodological and biomedical research by providing a rigorous testbed for assessing robustness, generalizability, and anatomical fidelity of generative models in heterogeneous multi-field MRI environments. By bridging ultra-low-field accessibility and ultra-high-field image quality within a single benchmark, this challenge is expected to catalyze the development of scalable, field-aware MRI synthesis methods and to support reliable multi-center neuroscience research and real-world clinical deployment of low-field MRI systems. The current release establishes a foundational harmonization benchmark, upon which pathology-aware extensions and clinically oriented downstream tasks will be systematically built in future editions.