A deep learning method for image-based subject-specific local SAR assessment.

Area of Science:

• Medical Imaging
• Computational Electromagnetics
• Artificial Intelligence in Medicine

Background:

• Local specific absorption rate (SAR) assessment is crucial for MRI safety but is typically measured via offline simulations using generic models, necessitating safety margins due to anatomical variability.
• Current methods lack patient-specific accuracy, leading to overestimation of SAR and potentially longer scan times.

Purpose of the Study:

• To develop and validate a deep learning-based method for image-based, subject-specific local SAR assessment.
• To train a convolutional neural network (CNN) to create a surrogate SAR model correlating subject-specific B1+ maps with local SAR.

Main Methods:

• A conditional generative adversarial network (cGAN) was trained using 5750 synthetic training samples derived from 23 subject-specific 7T body array models for prostate imaging.
• The network learned the relationship between complex B1+ maps and local SAR distributions, incorporating penalization for SAR underestimation errors.
• Both in silico and in vivo validations were performed to assess the method's accuracy and applicability.

Main Results:

• In silico cross-validation demonstrated good qualitative and quantitative agreement between predicted and ground-truth local SAR distributions.
• The peak local SAR estimation error showed a mean overestimation of 15% with a 13% probability of underestimation.
• In vivo validation confirmed the method's applicability with realistic data, showing good agreement with simulations.

Conclusions:

• The proposed deep learning method enables online, image-based, subject-specific local SAR assessment, significantly reducing uncertainty in current methods.
• This approach allows for less conservative safety factors and reduces MRI examination protocol time by approximately 25%.