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

This study investigates whether deep learning, specifically using Convolutional Neural Networks (CNNs), can improve the prediction of post-stroke aphasia severity compared to classical machine learning methods like Support Vector Machines (SVMs). The researchers hypothesized that CNNs, by modeling three-dimensional brain imaging data, would outperform SVMs and identify spatially dependent neuroanatomical information beyond the stroke lesion. Results showed that CNNs achieved higher accuracy and F1 scores than SVMs, even with nonlinear SVMs or dimensionality reduction techniques. Saliency maps revealed that CNNs leveraged distributed morphometry patterns across the brain, while SVMs focused on the lesion area. Ensemble clustering of CNN saliencies identified distinct morphometry patterns associated with aphasia severity, implicating unique networks linked to various cognitive processes. The study concludes that three-dimensional morphometry patterns are crucial for predicting aphasia severity and that CNNs offer a promising approach to improve outcome prognostication from neuroimaging data.

Alex Teghipco, Roger Newman-Norlund, Julius Fridriksson, Christopher Rorden, Leonardo Bonilha