A comparison of machine learning algorithms for automatic classification of neurons by their morphology

Abstract: Classification of neurons has been a studied topic in neuroscience for several years and with the increase of data, new methods are encouraged to help with the classification. This study compares different machine learning algorithms to see which are better suited for classification of large data sets of morphological reconstructions in multidimensional feature space. Ten algorithms were compared on a data set of over 10 000 samples of mice neurons. Further, each classifiers ability to classify each available cell type were also investigated. The results show that Random Forest had the best overall mean accuracy followed by Multi-layer Perceptron with 83% and 78% respectively. However, observing the classification of individual cell types all the algorithms varied in accuracy and Random Forest was not considered the best. In conclusion, machine learning algorithms are a viable source when classifying neurons but more research needs to be performed to reach the higher accuracy results.