Study On Whole-brain FMRI Classification Based On Generalized Sparse Logistic Regression

The brain is the material basis for higher neural activities such as consciousness,mind,language,learning,memory and intelligence in humans.Each human brain contains about 100 billion neurons and can generate more than one trillion pairs of connections between neurons.The human brain is the most subtle and complex system in the known universe.A variety of advanced imaging techniques have been applied to brain research,among which f MRI is widely used for its non-invasive,high resolution,and reproducible advantages.f MRI obtains three-dimensional brain imaging data by measuring the hemodynamic changes caused by neuronal activity.The classification of whole-brain f MRI data can effectively decode the cognitive state of the brain,which is important for understanding the working mechanism of the brain.Traditional machine learning algorithms such as support vector machines,linear discriminant analysis,kernel regression,and logistic regression are widely used in the classification of whole brain f MRI data.The dimensional catastrophe and overfitting problems faced by traditional algorithms in processing f MRI data can be avoided by introducing sparse penalty terms.However,the existing methods cannot fully utilize the spatial structure information of the brain,so there are limitations in directly using these methods for whole-brain f MRI data classification.To address this problem,this paper introduces both the penalty term characterizing sparsity and the penalty term characterizing spatial structure into the logistic regression algorithm,and proposes a generalized sparse logistic regression algorithm.In particular,a flexible penalty term characterizing the spatial structure is designed in this paper.The penalty term corresponds to a smoothing prior factor,which can make features with close spatial locations have similar weights,so as to achieve the smoothing effect of the features on the three-dimensional space.In the simplified one-dimensional model,the penalty term has a similar feature pattern to the Graph Net penalty term,which proves that the penalty term has the effect of feature smoothing.The parameter adjustment of this penalty term can fully utilize the spatial structure information of the brain to obtain higher classification accuracy and more neurophysiologically meaningful feature patterns.Under the framework of minorization-maximization,an iterative process is designed to solve the corresponding optimization problem,which guarantees that there is an explicit solution at each iteration step and eventually converges to the local optimal solution of the problem.On this basis,we introduce the idea of coordinate descent to construct a new iterative algorithm to further reduce the computational complexity of the algorithm.Experiments on simulated datasets show that the weight results obtained by the generalized sparse logistic regression algorithm have significant advantages in sparsity and smoothness relative to those obtained by the existing algorithms.Experiments on two standard f MRI datasets,namely the Star Plus dataset and the VOR dataset,show that the generalized sparse logistic regression algorithm has higher classification accuracy relative to the existing algorithms.