Distinct brain morphometry patterns revealed by deep learning improve prediction of post-stroke aphasia severity

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Imaging

Background:

• Post-stroke aphasia severity is linked to brain integrity beyond the lesion.
• Interindividual variability in aphasia severity remains unexplained by lesion anatomy alone.
• Three-dimensional morphometric patterns (e.g., atrophy) may account for this variability.

Purpose of the Study:

• To compare deep learning (CNNs) with classical machine learning (SVMs) for predicting chronic stroke aphasia severity.
• To evaluate if encoding spatial dependencies in whole-brain morphometry improves prediction accuracy.
• To identify unique predictive morphometric patterns beyond lesion location.

Main Methods:

• Convolutional Neural Networks (CNNs) applied to whole-brain morphometry and lesion anatomy.
• Comparison with Support Vector Machines (SVMs), including nonlinear and dimensionality reduction techniques.
• Analysis of saliency maps to interpret CNN predictions and identify predictive patterns.

Main Results:

• CNNs outperformed SVMs in predicting aphasia severity, achieving higher balanced accuracy and F1 scores.
• CNNs identified distributed morphometry patterns, while SVMs focused on the lesion area.
• Distinct 3D morphometry patterns, independent of lesion size, were associated with aphasia severity and implicated specific neural networks.

Conclusions:

• Three-dimensional morphometric network distributions are directly associated with aphasia severity.
• CNNs show potential for improving neuroimaging-based outcome prognostication in stroke.
• Interrogating spatial dependencies at multiple scales is beneficial for multivariate analysis in neuroimaging.