A deep learning approach for quantifying CT perfusion parameters in stroke.

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Computed tomography perfusion (CTP) imaging is crucial for acute ischemic stroke assessment.
• Conventional methods like singular value decomposition (SVD) face challenges with parameter estimation accuracy, including oscillations and underestimation.
• Global arterial input function (AIF) usage can lead to erroneous physiological parameter calculations.

Purpose of the Study:

• To develop an advanced method for accurate physiological parameter estimation from CTP images.
• To address limitations of traditional SVD methods in CTP analysis.
• To improve the diagnostic and therapeutic efficiency of CTP in stroke patients.

Main Methods:

• A Transformer-based neural network was developed to analyze voxel-wise temporal features in CTP images.
• The network utilized global AIF and brain tissue concentration time curves (CTC) as inputs.
• It estimated local AIF and the flow-scaled residue function, enabling calculation of cerebral blood flow (CBF) and bolus arrival delay (BAD).

Main Results:

• The proposed method demonstrated high accuracy in estimating local AIF (R=0.97 ± 0.04), CBF (error 4.95 ml/100 g/min), and BAD (error 0.51 s) on simulated data.
• Compared to SVD methods, the Transformer network significantly reduced estimation errors and avoided the 10-15% underestimation of CBF observed with SVD.
• Patient data analysis confirmed significantly higher CBF estimates with the new method in both normal and ischemic tissues.

Conclusions:

• The developed Transformer-based method accurately estimates local AIF and perfusion parameters from CTP data.
• This approach offers a significant improvement over conventional SVD methods for CTP analysis in acute ischemic stroke.
• The enhanced accuracy has the potential to improve CTP's role in diagnosing and managing ischemic stroke.