New Neural Field Method Enables 16x Faster Dynamic 3D MRI Scans
Breakthrough MRI reconstruction technique handles 16x undersampling while preserving motion detail.
Conventional MRI reconstruction treats images and coil sensitivities as discrete objects, creating memory bottlenecks and limiting regularization—especially in dynamic 3D cardiac MRI where undersampling is extreme. A new paper on arXiv proposes a model-based framework that replaces discrete representations with continuous, differentiable functions using tensor products of univariate neural fields. This tensor product structure scales efficiently to high-dimensional spatiotemporal scenarios, making it practical for real-world dynamic imaging.
The method outperforms state-of-the-art model-based reconstructions in both 2D and 3D dynamic settings. It preserves fine structure and motion even under aggressive undersampling (acceleration factor 16), a regime where traditional methods struggle. By eliminating discretization errors and reducing memory demands, this approach could enable faster, higher-quality cardiac MRI scans—reducing scan time without sacrificing diagnostic accuracy.
- Uses tensor product of univariate neural fields for continuous, discretization-free representation of magnetization and coil sensitivities.
- Achieves state-of-the-art reconstructions in dynamic 2D and 3D MRI at acceleration factor 16.
- Memory-efficient framework scalable to high-dimensional spatiotemporal settings like dynamic cardiac imaging.
Why It Matters
Could dramatically speed up cardiac MRI scans while maintaining image quality, improving patient comfort and clinical throughput.