Deep learning-driven inversion framework for shear modulus estimation in magnetic resonance elastography

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Science

Background:

• Magnetic Resonance Elastography (MRE) estimates tissue stiffness using inversion algorithms.
• The standard Multimodal Direct Inversion (MMDI) algorithm assumes ideal conditions and is sensitive to noise.
• There is a need for more robust and accurate MRE inversion techniques.

Purpose of the Study:

• To introduce a deep-learning-driven inversion framework for shear modulus estimation in MRE, named DIME.
• To improve the robustness and accuracy of MRE stiffness estimation compared to existing methods.
• To validate DIME's performance in simulations and in vivo human liver data.

Main Methods:

• DIME was trained using finite element modeling (FEM) generated displacement-stiffness data.
• The model utilized small image patches to capture local wave behavior and enhance robustness.
• Validation included homogeneous/heterogeneous FEM datasets, anatomy-informed liver simulations, and in vivo human liver MRE data.

Main Results:

• DIME produced stiffness maps with low variability and accurate boundaries in simulations, outperforming MMDI.
• In anatomy-informed simulations, DIME achieved high fidelity (r=0.99, R²=0.98) compared to ground truth.
• In vivo, DIME preserved physiological stiffness patterns, while MMDI showed systematic bias.

Conclusions:

• DIME demonstrates superior correlation with ground truth in simulations and comparable in vivo results to MMDI.
• MMDI's bias may stem from directional filtering, whereas DIME shows greater robustness.
• DIME is a feasible deep learning approach for clinical MRE applications.