Experimental Validation of Deep Learning Based Local SAR Mapping from 3D B1+ and Localizer-like MRIs for 7T pTx Systems

Local SAR related heating is a limiting safety factor at UHF MRI. Since local SAR cannot be measured, it should either be estimated from the global SAR or predicted by using the available data such as B1+ maps and MRIs. Here we proposed to use a multi-channel 3D CNN to utilize different channels for B1+ maps and MRIs simultaneously to improve local SAR prediction. We validated our method on 6 participants and found that maximum SAR values can be predicted with 80% accuracy, and local SAR variation can be predicted over 95% accuracy by using common image similarity metrics.