Machine learning segmentation of core and penumbra from acute stroke CT perfusion data

Area of Science:

• Neurology
• Radiology
• Medical Imaging

Background:

• Computed tomography perfusion (CTP) imaging is crucial for diagnosing acute ischemic stroke.
• Current methods rely on single perfusion thresholds, which can be imprecise.
• Accurate assessment of ischemic tissue is vital for treatment decisions.

Purpose of the Study:

• To develop and evaluate an alternative model for estimating tissue fate in acute ischemic stroke.
• To utilize machine learning with multiple CTP parameters for improved prediction.
• To move beyond traditional single-threshold limitations.

Main Methods:

• Employed machine learning (ML) models with four algorithms.
• Combined four CTP measures (CBF, CBV, MTT, delay time) and 3D-neighborhood analysis.
• Developed the model on 86 patient images and validated on 22.

Main Results:

• XGBoost demonstrated the highest performance among the tested algorithms.
• The ML model showed moderate agreement (Dice: 0.38-0.50) in segmenting core and penumbra compared to thresholding.
• Skull artifacts impacted accuracy, highlighting areas for improvement.

Conclusions:

• The developed ML model offers a promising alternative to traditional CTP thresholding for acute ischemic stroke assessment.
• Further refinement could enhance accuracy and overcome limitations of current methods.
• This approach may lead to more precise identification of salvageable brain tissue.