Comparison of classification methods for tissue outcome after ischaemic stroke

Area of Science:

• Neuroimaging
• Machine Learning
• Stroke Research

Background:

• Accurate identification of at-risk brain tissue in acute ischemic stroke is crucial for timely clinical intervention.
• Magnetic resonance imaging (MRI) data offers potential for classifying tissue at high risk of infarction.

Purpose of the Study:

• To evaluate and compare the performance of five popular classification methods for identifying infarction-risk tissue in acute ischemic stroke.
• To assess these methods using multimodal MRI data, including diffusion-weighted imaging and perfusion-weighted imaging.

Main Methods:

• Five classification algorithms (adaptive boosting, logistic regression, artificial neural networks, random forest, support vector machine) were applied to voxel-based brain imaging data from 55 acute ischemic stroke patients.
• Eight MRI parameters were utilized, encompassing diffusion-weighted imaging and perfusion-weighted imaging.
• Performance was assessed using area under the receiver operating curve (AUC_roc), area under the precision-recall curve (AUC_pr), sensitivity, specificity, and Dice coefficient.

Main Results:

• All methods demonstrated comparable sensitivity and specificity.
• Adaptive boosting, logistic regression, artificial neural networks, and random forest showed significantly better AUC_roc and Dice coefficient results.
• No statistically significant difference was observed among the five methods concerning AUC_pr, the primary performance metric.

Conclusions:

• Machine learning techniques, utilizing multimodal imaging data, can yield valuable prognostic information in acute ischemic stroke.
• These findings can support personalized treatment strategies for clinicians, pending thorough validation on independent datasets.