CT perfusion parameter estimation in stroke using neural network with transformer and physical model priors

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• CT perfusion (CTP) imaging is crucial for acute ischemic stroke management, enabling the identification of salvageable brain tissue and the infarcted core.
• Traditional CTP parameter estimation methods, like singular value decomposition (SVD), are susceptible to noise and arterial input function inaccuracies.
• Deep learning methods have not been previously applied to CT perfusion parameter estimation.

Purpose of the Study:

• To introduce CTPerformer-Net, a novel deep learning model based on the Transformer architecture for enhanced CT perfusion parameter estimation.
• To integrate physical priors into the deep learning model to improve the accuracy and robustness of perfusion parameter estimation.
• To evaluate the performance of CTPerformer-Net against traditional methods and commercial software.

Main Methods:

• Developed CTPerformer-Net, a Transformer-based deep learning model for CT perfusion parameter estimation.
• Incorporated physical consistency, smoothness, and physical model priors into the loss function design.
• Generated a simulation dataset using a physical model prior for training and validation.

Main Results:

• CTPerformer-Net demonstrated superior performance on the simulation dataset, with a 23.4% increase in correlation coefficients, a 95.2% decrease in system error, and a 90.7% reduction in random error compared to block-circulant SVD.
• The model successfully identified hypoperfused and infarcted lesions in 103 real CTP images from the ISLES 2018 challenge dataset.
• Achieved a mean dice score of 0.36 for infarct core segmentation, slightly outperforming commercially available software (dice coefficient: 0.34).

Conclusions:

• CTPerformer-Net significantly outperforms block-circulant SVD in CT perfusion parameter estimation, as evidenced by simulation data.
• The model's validity is confirmed through successful application on a real-world patient dataset.
• CTPerformer-Net represents a promising advancement in AI-driven CTP analysis for acute ischemic stroke.