ElastoNet: Neural network-based multicomponent MR elastography wave inversion with uncertainty quantification.

Area of Science:

• Medical Imaging
• Biophysics
• Computational Neuroscience

Background:

• Magnetic Resonance Elastography (MRE) is crucial for non-invasively measuring soft tissue stiffness.
• Traditional MRE inversion techniques struggle with noise and compression waves, limiting accuracy.
• Existing neural network approaches for MRE lack generalizability and uncertainty quantification.

Purpose of the Study:

• To introduce ElastoNet, a novel neural network for MRE wave inversion.
• To enable independent analysis of multiple wave components, improving generalizability across resolutions and frequencies.
• To provide uncertainty quantification maps for MRE parameter reconstruction.

Main Methods:

• ElastoNet utilizes evidential deep learning for uncertainty quantification.
• The network was trained on synthetic MRE wave patches (5x5 pixels).
• Evaluated on synthetic data, finite element simulations, phantom MRE, and human volunteer abdominal MRE studies (20-80 Hz).

Main Results:

• ElastoNet demonstrated comparable or superior accuracy in shear wave speed mapping compared to LFE, k-MDEV, and TWENN.
• Achieved lower root mean square error against ground truth in simulations and phantom data.
• Successfully generated uncertainty maps, a key advantage over existing methods.

Conclusions:

• ElastoNet offers a promising, generalizable solution for neural network-based MRE inversion.
• It effectively addresses noise and compression wave challenges in MRE parameter reconstruction.
• The method expands the potential of neural networks in diagnostic MRE.