Research On Feature Selection Method For FMRI Data Based On Regularized Sparse Models And Deep Neural Networks

Functional magnetic resonance imaging(f MRI) is an important brain imaging method in neuroimaging. It has been widely used in brain science and brain disease research because f MRI has non-invasive, repeated measurements and high resolution advantages. In recent years, the use of machine learning to classify brain cognitive states is an important research issue in bioinformatics based on f MRI data. Classifying the f MRI data can effectively decode brain stimuli, mental states and behaviours. It is very important to understand the working mechanism of the brain for human. However, the number of features far exceeds the number of data instances in f MRI data, which is a great challenge to this study. Training data instances is too small, which may lead to classification model overfitting. The dimension of the feature space is too high, which leads to high computational complexity and will also reduce the interpretability of the classification results simultaneously. In this paper, aiming at solving the problem of high dimensions and small samples in f MRI data, using regularized sparse models and deep neural networks to select features, we carry out the following two aspects of research:(1)To solve the problem of high dimensions in f MRI data, this paper proposes a feature selection framework of whole-brain f MRI data based on regularized softmax regression. Firstly, the whole brain is divided into region of interest(ROI) and region of non-interest(RONI). Then, L2-norm regularization is used for selecting out all voxels in ROI and L1-norm regularization is used for selecting out the activated voxels in RONI. L2-norm and L1-norm regularization all can shrink the size of coefficients, but L1-norm regularization can introduce a sparse solution. The voxels with non-zeros coefficients become selected voxels in prediction models. Finally, regularized softmax regression model of whole-brain f MRI data integrates the voxels in ROI and the voxels in RONI. The results in Haxby datasets show that L2-norm and L1-norm regularization strategy leads to more superior whole-brain classification performance compared to other existing methods.(2) To solve the problem of small samples in f MRI data, this paper proposes a feature selection method for f MRI data based on a deep stacked autoencoder neural network. In this method, we increase unlabelled f MRI data for unsupervised feature learning. Firstly, we employ all the unlabelled data and labeled training data to train the stacked autoencoder neural network for all the network parameters in autoencoder neural network and features in hidden layer. Then, Softmax classifier is trained with supervision by the features in the last hidden layer and the labels of training data for the parameters of softmax classifier. Finally, the value of model parameters in the pretraining phase can be used as initial value of model parameters in the finetuning phase. Back propagation algorithm is employed to fine tune the model. The results of f MRI data classification in Haxby datasets show that the proposed method leads to more superior performance compared to other existing methods using unlabelled data in resting state.