Stroke causes behavioral deficits in multiple cognitive domains and there is a growing interest in predicting patient performance from neuroimaging data using machine learning techniques. Here we investigated a deep learning approach based on Convolutional Neural Networks (CNNs) for predicting the severity of language disorder from 3D MRI lesion images from magnetic resonance imaging (MRI) in a heterogeneous sample of stroke patients. CNN performance was compared to that of conventional (shallow) machine learning methods, including Ridge Regression on the images’ principal components and Support Vector Regression. We also devised a hybrid method based on re-using CNN’s high-level features as additional input to the Ridge Regression model. Predictive accuracy of the four different methods was further investigated in relation to the size of the training set and the level of redundancy across lesion images in the dataset, which was evaluated in terms of location and topological properties of the lesions. The hybrid model achieved the best performance in most cases, thereby suggesting that the high-level features extracted by CNNs are complementary to PCA features and improve the model’s predictive accuracy. Moreover, our analyses indicate that both the size of training data and image redundancy are critical factors in determining the accuracy of a computational model in predicting behavioural outcome from the structural brain imaging data of stroke patients.
