FreeTune4D: Anatomy-Aware 4D-MRI Motion Reconstruction Benchmark and Free Fine-Tuning Framework.
Authors
Abstract
Deep learning and prior-image-guided cross modality motion reconstruction methods have recently en abled faster acquisition time and lower artifacts in abdominal four-dimensional magnetic resonance imaging (4D MRI). However, the limited availability and diversity of datasets and benchmarks significantly hinder generalizable model training and standardized evaluation, while insufficient anatomical structure awareness further limits performance on unseen datasets with varying pixel distributions. To address these challenges, we establish two large-scale, diverse, anatomically labeled digital phantom datasets: Digital4D-1M for training and Digital4D-900 for benchmarking. Specifically, Digital4D-1M contains one million clinically mimicking 4D-MRI samples, capturing variations in contrast, geometry, and quality across a 35-dimensional physiological parameter space, while Digital4D-900 ensures physiological consistency across three imaging modalities spanning three respiratory ranges. Building on these, we further propose FreeTune4D, a novel anatomy aware 4D-MRI motion reconstruction framework trained with a free fine-tuning strategy. This affine-to-deformable framework integrates vertebral stability during physiological motion and organ-level anatomical priors. As the first motion reconstruction method trained on large-scale digital phantoms, FreeTune4D surpasses state-of-the-art (SOTA) algorithms on Digital4D-900 in both affine alignment and deformable motion reconstruction. Additionally, in a real patient cohort of 133 cases (64 from Institution A, 16 from Institution B, and 53 from Institution C), FreeTune4D consistently outperforms SOTA methods across various metrics and modalities, demonstrating superior organ-level motion accuracy and structural consistency. These results highlight FreeTune4D's robustness and clinical potential for generalizable 4D-MRI motion reconstruction. The proposed datasets, code, as well as the model are available at https: //github.com/peilin-wang-git/FreeTune4D.