FMRI Data Classification Approach Based On Spiking Convolutional Neural Networks

With the continual advancement of neuroimaging technology,functional Magnetic Resonance Imaging(fMRI)has become a mainstream imaging technique that is safe,noninvasive,and useful for studying the working mechanism of the human brain as well as facilitating medical diagnosis.In recent years,there has been an increasing interest in combining brain neuroscience with machine learning,particularly in the effective processing and analysis of fMRI data.However,due to the high-dimensional,temporal,and spatial characteristics of fMRI data,significant challenges have arisen.The highdimensional nature of fMRI data makes it difficult for traditional machine learning methods to extract data features fully,and although some deep learning models,such as convolutional neural networks,have high modeling capabilities,they struggle to obtain good classification results due to the lack of full use of the temporal characteristics of fMRI.To address these issues,this paper proposes a spike train-based spiking convolutional neural network supervised learning algorithm through the linear approximation of the spike response function.The proposed method is applied to fMRI data classification and neurological disease auxiliary diagnosis system to accurately identify neurological diseases,providing patients with better treatment options,and improve the efficiency of diagnosis and treatment.The main research content of this paper includes three parts:(1)Aiming at the problems of non-differentiable spike function and gradient explosion that may exist in the learning process of traditional spiking neural network,two new spike response functions are proposed based on the spike response model: Triangle Linea-Postsynaptic potential(Triangle Linear-PSP,TL-PSP)function and Depolarization Linearized-Postsynaptic potential(Depolarization Linear-PSP,DL-PSP)function.At the same time,on the basis of the new spike response function,the Triangle Linear-Spiking Convolutional Neural Network(TL-SCNN)supervised learning algorithm and the Depolarization Linear-Spiking Convolutional Neural Network(DL-SCNN)were respectively proposed.Finally,the spiking convolutional neural network is applied to the CIFAR-10 dataset to verify the performance of the algorithm.Experiments show that compared with the traditional supervised learning algorithm of spike response model,the two algorithms proposed in this paper have higher recognition accuracy.In addition,compared with other spiking convolutional neural network learning algorithms,our algorithm still performs better.(2)Aiming at the high-dimensional,temporal and spatial characteristics of fMRI data,a spiking convolutional neural network fMRI classification model based on spike train was proposed and verified on the ADHD-200 dataset.Firstly,the fMRI data are preprocessed to extract brain regions of interest in fMRI images using common preprocessing techniques.Secondly,the fMRI images are encoded into spike trains using Step Forward encoding,and combined with the proposed TL-SCNN and DL-SCNN supervised learning algorithms for training.The two proposed models achieved accuracy rates of 80.2% and 78.7%,respectively.At the same time,the influence of model parameters on the classification performance was also analyzed.(3)The proposed spike convolutional neural network fMRI data classification model is used to construct an auxiliary diagnosis system for neurological diseases.The system has different users with varying responsibilities,including technicians who collect fMRI images,upload fMRI images,and train and upload fMRI training results,and attending doctors who judge whether the system’s results are correct through their clinical experience and correct the result.The system has significant application value for the intelligent diagnosis of neurological diseases.