Click here to

Session: Deep Learning Image Formation and Motion Management [Return to Session]

Physics-Informed Deep Learning for Accurate Material Density Map Generation Using MRI and DECT

C Chang1*, R Marants2, Y Gao3, M Goette4, J Scholey5, J Bradley6, T Liu7, J Zhou8, A Sudhyadhom9, X Yang10, (1) Emory University, Atlanta, GA, (2) Brigham and Women's Hospital, Boston, MA, (3) Emory University, Atlanta, GA,(4) Emory Healthcare, Atlanta, GA, (5) University of California San Francisco, San Francisco, CA, (6) Emory University School Of Medicine, Atlanta, GA,(7) Emory University, Atlanta, GA, (8) Emory University, Atlanta, GA, (9) Brigham and Women's Hospital | Dana-Farber Cancer Institute | Harvard Medical School, Boston, MA, (10) Emory University, Atlanta, GA


WE-C930-IePD-F5-5 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: This work explores the feasibility of using physics-informed deep learning (PIDL) to integrate MRI and dual-energy CT (DECT) information to enhance material mass density inference for proton Monte Carlo dose calculation.

Methods: We propose a PIDL-based multimodal imaging framework to correlate MRI and DECT to material mass density constrained by a physics model. We evaluated three models: an empirical model and physics-informed ResNet-DECT (PRN-DE) using DECT data and physics-informed ResNet-MR-DECT (PRN-MR-DE) trained by DECT and MRI data. Five tissue substitute MRI phantoms were used for deep-learning-based material calibration including adipose, skin, muscle, liver, and cortical bone. The training inputs include T1-weighted Dixon water- and fat-only images, T2-weighted STIR, twin-beam images acquired at 120 kVp with gold (Au) and tin (Sn) filters, virtual monochromatic images (VMI) of 80 keV, and relative electron density map. The reference mass densities of each phantom were based on relative stopping powers measured by a 200 MeV proton spot. A retrospective patient study included a head-and-neck patient to conduct a proof-of-concept test of the proposed method since the ground truth mass density map was not available.

Results: The empirical model, PRN-DE, and PRN-MRDE result in mass density errors of: 1.4%±4.0%, 2.9%±2.0%, and 0.4%±3.9% for adipose; 1.0%±0.8%, 1.6%±0.8%, and 0.5%±0.9% for skin; 0.1%±0.9%, -0.6%±1.0%, and 0.0%±0.9% for muscle; -0.3±6.9%, -0.5±3.0%, and 0.3±3.0% for liver; -0.7%±3.8%, -0.8%±1.4%, and -0.4%±1.2% for cortical bone. For patient study, the density predicted by PRN-MR-DE shows consistent trend compared to Hounsfield unit (HU) variations of 80-keV VMI, while PRN-DE results in large deviations.

Conclusion: The proposed PIDL-based multimodal imaging framework demonstrated that more accurate material mass density conversion could be achieved by ResNet using MRI and DECT images than ResNet using only DECT images. The patient study shows that PRN-MR-DE has the potential to enhance proton treatment uncertainty due to material conversion.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email