Deep Learning-Based Prediction of Specific Absorption Rate Induced by Ultra-High-Field MRI RF Head Coil

Objective: As magnetic resonance imaging (MRI) technologies advance, predicting local Specific Absorption Rate (SAR) distributions becomes more challenging. This difficulty arises from the unique anatomical structures and dielectric properties of individual subjects, coupled with the inherent non-uniformity of energy deposition within tissues during scanning. To rapidly estimate SAR values induced by ultra-high-field (UHF) MRI birdcage RF coil in near real-time, this paper proposes a deep learning-based framework. Methods: The proposed framework consists of two stages. During the dataset generation stage, high-dimensional model representation, a polynomial-based surrogate modeling technique, is used to generate a large and diverse dataset, thereby reducing the reliance on resource-intensive deterministic simulations performed by physics-based simulators. During the inference stage, the framework employs 3D Attention U-Net, processing relative permittivity and conductivity maps of head models along with incident electric fields to predict SAR distributions. Results: The 3D Attention U-Net outperforms all other 3D U-Net variants and demonstrates remarkable accuracy, with mean relative errors of 7.57% for voxel SAR, 5.63% for 10g-averaged SAR, and 2.60% for peak spatial SAR. Each prediction can be performed in less than half a second, outperforming traditional physics-based simulators by at least three orders of magnitude. Conclusion: The framework provides a significant computational advantage over traditional physics-based simulators while maintaining satisfactory accuracy. Significance: The computational framework, available on GitHub, enables real-time SAR predictions on permittivity and conductivity distributions on any unseen MRI head models. The framework will allow ultra-fast optimization and uncertainty quantification studies to be performed while designing new UHF MRI coils.