Model-Agnostic Deep Learning Approaches for Dynamic Magnetic Resonance Thermometry Reconstruction

Motivation: Effective real-time temperature monitoring is essential in FUS therapies, but existing MR thermometry methods often sacifice phase accuracy and speed, limiting clinical applications. Goal(s): This study aims to enhance both accuracy and real-time capability in MR thermometry by optimizing a deep learning-based reconstruction method. Approach: A residual UNet model with four optimizations-offline and online data augmentation, knowledge distillation, and amplitude-phase decoupled loss, was applied to phantom, ex vivo, and clinical datasets, with 2× and 4× k-space under-sampling. Results: The optimized ResUNet model demonstrated superior temperature map accuracy and speed, achieving effective acceleration rates of 1.9 and 3.7 for 2× and 4× under-sampling, respectively. Impact: The study's results enable faster and more accurate MR thermometry, allowing clinicians to better monitor and control temperature in thermal therapies. This advancement encourages further investigation into deep learning's role in MR applications, potentially extending to other phase-sensitive imaging techniques.