Experimental Validation of Subject-Specific Local SAR Assessment by Deep Learning

Subject-specific local SAR assessment is one of the outstanding challenges of ultra-high field MRI. In this work, we present experimental, in-vivo results on four healthy subjects of a novel deep learning approach which provides online subject-specific local SAR distributions based on measured B₁⁺ maps. Results are validated by creating a subject-specific model of each subject and calculate a reference local SAR distribution off-line by FDTD simulations. The results show that a Convolutional Neural Network (CNN) trained with synthetic MR data enables accurate subject-specific local SAR prediction during an MRI examination.