Deep-Learning based Referenceless MR Thermometry Using Residual U-Net with Self-Attention

Motivation: Clinical use of referenceless MR thermometry is challenged by inconsistent spatial patterns of phase distribution, which limit the accuracy of temperature measurements for motion objects. Goal(s): This study aims to develop a robust deep-learning model for reconstructing absent phase information on focal regions, regardless of heating spot location, eliminating dependence on specific functions. Approach: A residual U-Net with a self-attention mechanism was used to restore the background phase based on full-phase data from the subject region. Results: The model achieved high coherence with ground truth temperatures, particularly when focal areas were adjacent to boundaries where spatial phases are complicated, confirming reliable referenceless thermometry. Impact: This work enhances MR thermometry in organs with respiratory motion by achieving accurate, real-time temperature measurements and overcoming regional phase variability, demonstrating its potential for broader clinical applications in dynamic thermal assessment.