Research On Feature Extraction And Classification Of Neurons Based On Multivariate Statistical Analysis

One of the major events in the 30 years development of natural science is neuroscience and brain science rapid rise,more and more facts proved that neuroscience may lead to the rapid development in the life science of another climax since the 21 st century. The thesis studies neuron classification, selecting the data of the neuron geometrical features as the research object, which uses multivariate statistical method.The main data in this study derives from NeuroMorpho. Org. The original data of 62 neurons downloaded is described in standard SWC file format. Moreover, it uses software tools L-Measure extract 43 geometry characteristics of the neurons from the original data and selects a research significance from seven indices of each feature, or analyzes a index combining several indices of great research significance weighted. In addition, it also constructs the discrete coefficient screening criteria and selects 27 geometric features as well as reduces the dimensionality of 27 geometric features Using the method of fact analysis, which eventually transforms into six factors, including:(1) neurons close situation,(2) neuronal size,(3) neurons of the bifurcation,(4) motor neurons(5) branch and branch of neurons,(6) neurons branch points.According to six comprehensive factors of each neuron, application of cluster analysis classifies neurons, finding out clustering result by feature selection is in accordance with classification results in terms of neuronal function while accuracy of clustering result without feature selection is relatively low and the hierarchical graph is not easy to distinguish categories of neurons. As a result, we can see clustering result with feature selection is superior to that without feature selection through analysis and comparison.In terms of six comprehensive factors of each neuron, applying discriminant analysis to study, choosing 70 percent of neurons as training samples, discriminant function corresponding seven types of neurons is obtained. The effect turns to be good, though there is still a error ratio at 0.053 by inspecting 30 percent of testing samples.