A spatiotemporal dependency-aware lightweight CNN-ViT network for 3D MRF with a balanced acceleration strategy.
Authors
Affiliations (6)
Affiliations (6)
- College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, China.
- School of Communication Engineering, Hangzhou Dianzi University, Hangzhou, China. Electronic address: [email protected].
- GE Healthcare, MR Research, Beijing, PR China.
- Department of Radiology, Stanford University, Stanford, CA, United States; Department of Electrical Engineering, Stanford University, Stanford, CA, United States.
- Department of Imaging Sciences, University of Rochester, Rochester, NY, United States.
- School of Physics, Zhejiang University, Hangzhou, China; State Key Laboratory of Brain-Machine Intelligence, Zhejiang University, Hangzhou, China. Electronic address: [email protected].
Abstract
The push for rapid MRI acquisition aims to enhance clinical efficiency and diagnostic consistency by shortening scan times. 3D Magnetic Resonance Fingerprinting (MRF) has emerged as a promising technique for fast, multi-parametric quantitative imaging. However, its accuracy and relatively long acquisition time remain a limiting factor for clinical adoption. Accelerating MRF while preserving quantitative accuracy constitutes a crucial research objective. Deep learning approaches have recently been applied to accelerate MRF parameter quantification, but existing methods still exhibit notable limitations in both acceleration scheme design and the ability to model the complex contextual information embedded in MRF data. To address these limitations, we propose a lightweight spatiotemporal attention enhanced network (LiST-UNet) that integrates convolutional neural networks with lightweight Vision Transformer components to model long-range spatiotemporal dependencies in 3D MRF. A precursor-successor network is included to model interrelationships among tissue parameters, improving T2 quantification accuracy, while a balanced k-space and temporal-frame acceleration strategy significantly reduces errors compared with single-dimension undersampling schemes. Experimental results demonstrate that the proposed method enables whole-brain MRF imaging in approximately 1.25 min, achieving an eightfold acceleration over conventional 10-minute acquisitions with superior quantification accuracy and image quality compared to previously proposed deep learning methods. This work combines architectural improvements in MRF reconstruction with an acceleration strategy, supporting the future clinical translation of 3D MRF.