Decoding post-stroke motor function from structural brain imaging.

Area of Science:

• Neuroimaging
• Machine Learning
• Stroke Research

Background:

• Machine learning (ML) enhances clinical neuroimaging by enabling individualized predictions and analyzing complex brain interactions.
• Applying ML to structural imaging for brain injury, like stroke, is challenging due to high lesion variability.
• Extracting meaningful features from anatomical images for ML remains an open research question.

Purpose of the Study:

• To compare two feature extraction methods for stroke lesion data: lesion load per region versus voxel patterns.
• To evaluate different anatomical area delimitations for feature extraction in stroke patients.
• To determine the optimal methodology for predicting long-term motor outcomes using early post-stroke structural MRI.

Main Methods:

• Employed Gaussian Process Regression (GPR) on structural MRI data from 50 chronic stroke patients.
• Compared lesion load per region against voxel pattern feature extraction methods.
• Investigated various anatomical area definitions: atlas-based regions, corticospinal tract, fMRI-derived masks, and lesion-symptom mapping regions.

Main Results:

• Feature extraction using voxel patterns representing lesion probability yielded superior results compared to lesion load per region.
• The best predictive performance was achieved using a combined set of cortical and subcortical motor areas, along with the corticospinal tract.
• This indicates the significance of detailed spatial lesion information over aggregated regional counts.

Conclusions:

• Voxel pattern-based feature extraction is more effective than lesion load for predicting motor outcomes in stroke patients.
• Combining specific motor-related brain regions and the corticospinal tract optimizes prediction accuracy.
• These findings provide crucial methodological guidance for leveraging early structural MRI in predicting stroke recovery.