Research On FMRI Data Classification Method Based On Neural Networks

With the development of neuroimaging technology,functional Magnetic Resonance Imaging(fMRI)is widely used in research on cognitive neuroscience and brain diseases.Taskevoked fMRI reflects the change in related functional activity of brain under the specific task stimulation,while resting-state fMRI reflects brain functional activity at resting state.The classification of fMRI data is meaningful,since it can enhance human study on decoding of brain status on one hand,and is conducive to the diagnosis and treatment of brain diseases on the other.In this paper,we study the classification of task-evoked fMRI data and resting-state fMRI data based on neural networks.In allusion to the characteristics of task-evoked fMRI data,we propose a novel parallel multi-channel two-dimensional convolutional neural networks(P2D CNN)for classifying the 3D brain images of task-evoked fMRI.This model preserves information integrity of whole brain voxel-wise features by transforming the 3D brain image into three multi-channel 2D images,and improves the ability of extracting features from 3D spatial information by using three multi-channel 2D convolutional neural networks in parallel.This model mainly uses 2D convolution method,which improve the efficiency of model training compared to 3D convolution method.Experiment results on public task-evoked fMRI dataset of HCP demonstrate that P2 D CNN outperforms other comparison models in classification and achieves high training efficiency.The functional connectome feature is an important and commonly used feature in the resting-state fMRI data classification.For its lack of other information of fMRI,we consider to add features of brain regions,whole brain voxels and personal characteristics in comprehensive analysis,and propose a model(CDNN,Convolutional and Deep Neural Networks)consisting of a 3D convolutional and a fully connected neural networks that can comprehensively combine these features for classification.This model learns the high-order features of whole brain voxelwise features through a 3D CNN and combines it with other feature by fully connected layer,which makes it take into account a variety of information for classification.We conduct classification experiment of autism patients using public resting-state fMRI dataset of ABIDE.Results show that CDNN has advantages in classification.At last the experiment discusses the influence of output dimension of high-order features on classification effect in CDNN.