High-Speed Tissue Quantification Using Spatiotemporal Attention Mechanism with Undersampling Strategy Consideration

Motivation: Current deep learning methods for accelerating MRF rely on extreme undersampling in k-space or time frames, but this is suboptimal. And the model designs struggle to fully utilize the spatiotemporal features of the data. Goal(s): To develop a method that identifies the optimal two-dimensional undersampling combination, surpassing single-dimension undersampling in quantitative performance. Approach: We enhanced Stolk et al.'s error model and developed a Spatiotemporal Attention with Tissue Specificity Network (STA-TS) that integrates tissue features across spatial and temporal dimensions. Results: This approach achieves 1mm full brain coverage in 2.5 minutes, with significantly improved quantitative accuracy for T1 and T2 compared to single-dimension undersampling. Impact: The proposed STA-TS network significantly improves MRF accuracy and efficiency, achieving full brain coverage at 1 mm resolution in just 2.5 minutes, providing a faster and more accurate imaging solution for clinical applications.