Physics Model-Aware MR Relaxometry Using Deep Learning
S Zhang, Y Gao, Y Zhong, C Shen, Y Peng, J Grandinetti, J Deng, X Jia*, The University of Texas Southwestern Medical Center, Dallas, TX
Presentations
WE-A-201-5 (Wednesday, 7/13/2022) 7:30 AM - 8:30 AM [Eastern Time (GMT-4)]
Room 201
Purpose: MR relaxometry delivering T1, T2, T2* maps provides quantitative information of clinical importance. Conventional methods measuring these maps require extensive scans with specific protocols. This study develops a deep learning-based method that incorporates physics models of MR signals to obtain quantitative relaxometry maps from conventional T1-weighted (T1w) and T2-weighted (T2w) images.
Methods: A two-channel convolutional neural network was trained to output T1, T2 and T2* values of each pixel based on T1w and T2w image patches centering at that pixel. Loss function was designed to enforce consistency indicated by paired input and output values in the training data, as well as the consistency among them governed by MR physics models. We tested the model in simulation and experimental cases. Simulated T1w and T2w images of a human brain were generated using a SPGR sequence and a SE sequence based on known T1, T2, T2* and proton density maps. In the experimental study, a mouse was scanned on a 3T scanner to obtain T1w and T2w images using a multi-echo FLASH and a SE sequence, inversion recovery SE images for ground truth T1 mapping, multi-echo SE images for T2 mapping, and multi-echo FLASH images for T2* mapping.
Results: The estimated T1, T2 and T2* maps for both the simulated and experimental data visually agreed with the ground truth maps. Quantitatively for the simulated data, the peak signal-to-noise ratios (PSNR) of T1, T2 and T2* maps in the white matter and gray matter were > 31, and the corresponding mean absolute percentage errors (MAPE) were all within 8%. For the experimental data, the PSNR was 19.5-33.5 and MAPE was 6.4-12.7%.
Conclusion: The proposed physics model-aware MR relaxometry model can provide T1, T2 and T2* maps using only conventional contrast-weighted images, potentially offering clinically useful information.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH R01CA22728901 and R01CA214639.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email