A physics-informed deep learning framework for dynamic susceptibility contrast perfusion MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Perfusion MRI is crucial for diagnosing and monitoring neurovascular and neurooncological diseases.
• Conventional processing methods for perfusion MRI lack standardization and struggle to capture detailed perfusion dynamics.

Purpose of the Study:

• To introduce a physics-informed deep learning framework for analyzing dynamic susceptibility contrast perfusion MRI data.
• To achieve accurate recovery of the dynamic tissue response using this novel framework.

Main Methods:

• Utilizes physics-informed neural networks (PINNs) to learn voxel-wise tissue residue functions (TRF).
• Employs total variation and elastic net regularization for stable network output.
• Calculates normalized cerebral blood flow (nCBF) and volume (nCBV) parameter maps from predicted residue functions.

Main Results:

• PINN-derived residue functions show high concordance with true functions in simulations.
• Calculated nCBF and nCBV values converge to true values with increasing contrast-to-noise ratio.
• High correlations and image similarity indices observed for nCBF and nCBV in an in vivo patient dataset.

Conclusions:

• Physics-informed neural networks (PINNs) offer a robust method for analyzing perfusion MRI data.
• PINNs can accurately and stably recover local vasculature response functions.
• This approach enhances the analysis of perfusion dynamics in neuroimaging.