Abstract
Purpose :
It is speculated that even the most complex cognitive function of the brain, such as attention, consciousness and decision-making, can be understood with the help of neuronal morphology. Morphologically distinct neuronal cell types exist at functionally divided brain regions. Therefore, the morphological analysis of neurons is of great importance to understand the brain function. However, given the complication of neuron morphology in the central nervous system, classifying these neurons quickly and accurately is challenging. It is crucial to develop an automatic analysis method. Using machine learning algorithms, we developed an ensemble algorithm to classify neurons with the accuracy of > 95%.
Methods :
84 morphological features were extracted from the histological digital reconstructed neurons of the layer 2/3 of the mice primary visual cortex. The features include soma size, dendrite density, axon length, axon centroid, etc. The parameters were then normalized to eliminate the influence of data amplitude. To minimize the influence of similar features during classification, we orthogonalized all morphological characteristics with the principal component analysis (PCA). On the basis of dimensionality reduction, we selected the first 30 normalized PCA eigenvectors to classify neurons, and compared several supervised machine learning algorithms, including k-nearest neighbor algorithm, logistic regression, linear support vector machine, nonlinear kernel function support vector machine, gaussian process, perceptron neural network, decision tree, random forest, Adaboost, naïve Bayes and quadratic discriminant analysis. We selected four classifiers and established an ensemble method.
Results :
All 11 classifiers (k-nearest neighbor algorithm, logistic regression, linear support vector machine, nonlinear kernel function support vector machine, gaussian process, perceptron, decision tree, random forest, Adaboost, naïve Bayes and quadratic discriminant analysis) were evaluated with 5-fold cross-validation. K-nearest algorithm, linear support vector machine, perceptron and naïve Bayes performed better than the others. The ensemble method showed high accuracy (95%) in classification of interneurons in L2/3 primary visual cortex.
Conclusions :
The ensemble machine learning method we developed has high accuracy (>95%). It is a convenient tool to reduce human labor and bias during neuronal morphological classification.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.