Use of a convolutional neural network to identify infarct core using computed tomography perfusion parameters

Area of Science:

• Neurology
• Radiology
• Medical Imaging

Background:

• Ischemic stroke diagnosis relies on computed tomography perfusion (CTP) imaging.
• Contralateral hemisphere comparisons of CTP parameters are standard but face challenges due to varying software and patient hemodynamics.
• Existing methods often require non-universal thresholds for accurate infarct tissue segmentation.

Purpose of the Study:

• To evaluate the efficacy of a convolutional neural network (CNN) for segmenting infarct tissue in ischemic stroke.
• To determine the most accurate CTP parameter for infarct segmentation using a CNN approach.
• To eliminate the need for non-universal CTP parameter thresholds in stroke imaging.

Main Methods:

• Retrospective analysis of CTP data from 63 ischemic stroke patients.
• Generation of CTP parameter maps (CBF, CBV, TTP, MTT, delay time) using Vitrea CTP software.
• Training and testing a U-net based CNN on 8,352 infarct slices, with Monte Carlo cross-validation, comparing CTP parameters against diffusion-weighted imaging (DWI) ground truth.

Main Results:

• Cerebral blood flow (CBF) demonstrated the highest spatial agreement (Dice coefficient=0.67, PPV=0.76) in predicting infarct volumes.
• CBF showed the smallest volume difference (14.3±11.5 mL) compared to DWI infarct volumes.
• Other CTP parameters like CBV, TTP, MTT, and delay time showed lower agreement and larger volume discrepancies.

Conclusions:

• Cerebral blood flow (CBF) is the most accurate CTP parameter for infarct segmentation.
• CNN-based infarct segmentation offers a promising method to overcome limitations of non-universal CTP thresholds.
• This approach could standardize infarct volume assessment in ischemic stroke patients.