Machine Learning Algorithms Versus Thresholding to Segment Ischemic Regions in Patients With Acute Ischemic Stroke

Area of Science:

• Neuroradiology
• Medical Imaging Analysis
• Machine Learning in Medicine

Background:

• Computed tomography (CT) scans and CT perfusion (CTP) are crucial for early stroke assessment.
• CTP parametric maps help differentiate infarct core and penumbra, guiding treatment.
• Automated segmentation of these regions is essential for efficient clinical workflows.

Purpose of the Study:

• To compare fully-automated machine learning and thresholding methods for segmenting hypoperfused regions in ischemic stroke patients.
• To evaluate the performance of different machine learning algorithms and architectures on CTP data.
• To determine the most effective automated approach for stroke lesion segmentation.

Main Methods:

• Utilized parametric maps from CTP datasets as input features.
• Compared two distinct machine learning architectures with three mainstream algorithms.
• Employed manual annotations by expert neuroradiologists as the ground truth for validation.

Main Results:

• The Random Forest (RF) algorithm achieved the highest performance.
• RF demonstrated an average Dice coefficient of 0.68 for penumbra and 0.26 for core.
• Volume differences were 25.1 ml for penumbra and 7.8 ml for core, indicating good accuracy.

Conclusions:

• The best-performing Random Forest-based method significantly outperformed classical thresholding approaches.
• This automated method accurately segments ischemic regions across varying stroke severities.
• Improved segmentation accuracy aids in better clinical treatment decisions for stroke patients.