ABSTRACT Purpose To enable the rapid generation of subject‐specific whole‐body anatomical models for patient‐specific prediction of torso–local specific absorption rate (SAR) in MRI. Methods A 6‐s 3D gradient‐echo MRI sequence was used to acquire data within the imaging field of view (FOV). Major tissue types were automatically segmented using a deep learning model trained via a semi‐supervised strategy combining teacher–student learning and partial‐category annotations. A full‐body geometry was reconstructed from depth data captured by a 3D camera, thereby extending the model beyond the FOV. The MRI‐derived anatomical segmentation and camera‐based external geometry were co‐registered and fused into a seamless, subject‐specific human model. Results Human models were generated in approximately 20 s per subject, including MRI acquisition and processing. Accurate tissue segmentation and robust body reconstruction were achieved. Validation on the Duke numerical phantom yielded an average peak SAR 10g error < 2%. In vivo field comparisons in 20 volunteers showed a normalized root‐mean‐square error (NRMSE) of 9.50%. The models preserved subject‐specific anatomy and were suitable for electromagnetic simulation. Conclusion A hybrid framework integrating ultrafast MRI, depth data scanning and deep learning enables rapid construction of subject‐specific human models, supporting practical, online SAR monitoring in clinical MRI.
Hu et al. (2026) studied this question.