Neurons Classification And Identification Based On Spectral Decomposition

Neuron is the basic structural unit in the human brain structure，has the characteristicsof variety relative to the identify category，complex structure relative to the sample ischaracterized by high dimension，etc．The classification of the neuron has always been aworldwide difficult problem in the plan of the human brain．If neurons can be classifiedand recognized correctly，it will be of great significance in the research and treatment ofnerve pathology．In the study of classification of neurons，the majority of researches are inmorphological characteristic，however，the variety and multi-feature of neurons cause greatdifficulties for classification．Support vector machine(SVM) algorithm can deal with multi-classification problem for high-dimensional data，and have better effect．But support vectormachine has not very good effect for classification of non-convex or irregular sampleset．So this paper made the following research work：(1) Adopting the method of spectrum decomposition to screen the characteristics ofneurons．Considering that spectral decomposition method is just suitable for two or three kindsof samples and SVM is a quadratic programming problem，according to the principles ofmaximum separation，this paper use the shortest path algorithm to combine multipleclasses into two kinds，which can make the sample set be a non-convex problem，thenscreen characteristics by the spectral decomposition that do not consider whether thesample distribution is convex model．(2) Using redundant weighted FCM algorithm to realize a coarse classification．In order to make the sample be effectively divided，using sparse technology to reducethe interference of some samples.Then using redundant weighted FCM algorithm toachieve coarse classification of samples．(3) Using local support vector machine to realize the refinement of sample points．Finally，20morphological feature of231neurons of4types are selected to be used inneuron classification in this paper，and adopting the method of the research in this paper totrain and recognize the sample of neuron．The experimental results show that the proposedapproach is feasible in the classification of neurons，and has a better effect，up to94.92%．Through experimental research，it is found that training samples selected and thechoice of parameters in the algorithm will be affect the result of the classification ofneurons to a certain extent．...