Identification of infarct core and ischemic penumbra using computed tomography perfusion and deep learning

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate assessment of infarct and penumbra volume is critical for acute ischemic stroke (AIS) management.
• Current CT perfusion (CTP) software relies on relative thresholding, a method with limitations and scientific debate.
• Novel approaches are needed to improve the precision of CTP in quantifying ischemic tissue.

Purpose of the Study:

• To investigate the efficacy of deep learning algorithms for precise infarct and penumbra segmentation on CTP hemodynamic maps.
• To compare the performance of deep learning models against commercial CTP software in AIS patients.
• To enhance the reliability of CTP in guiding clinical treatment decisions for AIS.

Main Methods:

• Retrospective collection of CTP scans from 119 AIS patients.
• Generation of cerebral blood flow, cerebral blood volume, and other hemodynamic maps using commercial CTP software.
• Training U-Net-shaped deep learning architectures for infarct and infarct+penumbra segmentation, employing test-time-augmentation, ensembling, and watershed segmentation for postprocessing.

Main Results:

• Deep learning algorithms achieved a Dice coefficient (DC) of 0.63-0.65 and mean absolute volume error (MAVE) of 4.5-5.32 mL for infarct segmentation.
• Commercial software showed significantly lower performance with DC of 0.26-0.36 and MAVE of 7.12-12.42 mL for infarct.
• For infarct+penumbra, deep learning achieved DC of 0.60-0.63 and MAVE of 5.91-7.11 mL, outperforming commercial software (DC 0.25-0.35, MAVE 7.25-11.11 mL).

Conclusions:

• Deep learning algorithms offer a precise method for assessing infarct and penumbra volumes in AIS.
• These algorithms significantly outperform existing relative thresholding methods used in commercial CTP software.
• Implementing such deep learning approaches can enhance the diagnostic accuracy and reliability of CTP for AIS treatment guidance.