Redes neuronales convolucionales para la clasificación de componentes independientes de rs-fMRI

Resting state functional magnetic resonance imaging (rs-fMRI) is one of the most relevant techniques in brain exploration. However, it is susceptible to many external factors that can occlude the signal of interest. In this order of ideas, rs-fMRI images have been studied adopting different approaches, with a particular interest in artifact removal techniques through Independent Component Analysis (ICA). Such an approach is a powerful tool for blind source separation, where elements associated with noise can be eliminated. Nevertheless, such removal is subject to the identification or classification of the components provided by the ICA. In that sense, this study focuses on finding an alternative strategy to classify the independent components. The problem was addressed in two stages. In the first one, the components (3D volumes) were reduced to images by Principal Component Analysis (PCA) and by obtaining the median planes. The methods achieved a reduction of up to two orders of magnitude in the weight of the data size, and they were shown to preserve the spatial characteristics of the independent components. In the second stage, the reductions were used to train six models of convolutional neural networks. The networks analyzed in this study reached accuracies around 98 % in classification, one of them even up to 98.82 %, which reflects the high discrimination capacity of convolutional neural networks.