Detecting cognitive states from fMRI images by machine learning and multivariate classification

The major obstacle in building classifiers that robustly detect a particular cognitive state across different subjects using fMRI images has been the high inter-subject functional variability in brain activation patterns. To overcome this obstacle, firstly, the brain regions that are relevant to the problem under study are determined from the training data; then, statistical information of each brain region is extracted to form regional features, which are robust to inter-subject functional variations within the brain region; finally, the regional feature statistical variations across different samples are further alleviated by a PCA technique. To improve the generalization ability and efficiency of the classification, from the extracted regional features, a hybrid feature selection method is utilized to select the most discriminative features, which are used to train a SVM classifier for decoding brain states from fMRI images. The performance of this method is validated in a deception fMRI study. The proposed method yielded better results compared to other commonly used fMRI image classification methods.