Magnetic Resonance Electrical Property Tomography (MREPT) could provide important contrast for non-calcified tumors. However, MREPT relies on numerical differentiation, which is noise sensitive and prone to artifacts near boundaries.
In this work, physics-informed neural networks (PINN), NN empowered automatic differentiation is proposed to improve MREPT by mitigating artifacts and reducing noise sensitivity. Instead of calculating partial derivatives numerically, is obtained by backpropagation through PINNs.
For clinical MREPT, reduction of ground-truth information to guide PINNs was investigated. Results show that above 25% collocation points, reconstruction can be made at 100 SNR. PINNs enable noise-robust and artifact-free MREPT from less ground-truth information.
This abstract and the presentation materials are available to members only; a login is required.