Spatio-temporal physics-informed learning: A novel approach to CT perfusion analysis in acute ischemic stroke

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• CT perfusion imaging is crucial for evaluating brain tissue in acute ischemic stroke.
• Standard CT perfusion analysis software struggles with noisy data, impacting accuracy.
• Software tuning is often required to mitigate noise effects in perfusion analysis.

Purpose of the Study:

• To develop a noise-robust CT perfusion analysis method using physics-informed learning.
• To introduce SPPINN (Spatio-temporal Perfusion Physics-Informed Neural Network) for enhanced cerebral perfusion estimation.
• To validate the performance of SPPINN on simulated and real-world patient data.

Main Methods:

• Proposed SPPINN, a spatio-temporal physics-informed neural network approach.
• Utilized implicit neural representations of contrast attenuation from CT perfusion scans.
• Validated on simulated data, in-house patient data, and public benchmark datasets.

Main Results:

• SPPINN achieved accurate perfusion parameter estimates despite high noise levels.
• The method successfully differentiated healthy brain tissue from infarcted tissue.
• SPPINN-generated perfusion maps showed strong correspondence with reference infarct core segmentations.

Conclusions:

• Physics-informed learning offers a noise-robust framework for CT perfusion analysis.
• SPPINN provides accurate cerebral perfusion estimation, even with noisy data.
• This approach enhances the reliability of imaging in acute ischemic stroke evaluation.