Improving Visual Neuroscience Cell Type Classification with Supervised Machine Learning*

At present, visual neuroscientists must employ an inefficient, time-intensive process to study the ways in which various types of neurons react to characteristics of a visual stimulus; the standard procedure requires specifying and monitoring a single cell type per individual microscopy recording. This research paper proposes an alternative method: utilize a supervised classification algorithm to distinguish between several cell types – based on the cells’ behavior and response to stimuli – in the context of a single recording. This allows researchers to record multiple cell types at once and, subsequently, classify them by type for further analysis. For this classifier, the neuronal spatial footprints and neuronal temporal activity are extracted from raw microscopy recordings using constrained nonnegative matrix factorization. From these data, neuronal features are engineered for the classifier, which-along with features engineered from the visual stimulus corresponding to the neuronal activity-are used by various models to predict the cell type of the recorded neurons. Several algorithms are tested to compare their classification performance, including random forest classifiers, neural networks, and K-nearest neighbors classifiers. This research concludes that the relationship between stimulus and fluorescent response is a moderate predictor of cell type. We develop a cell type classification model that leverages one-hot encoding and engineering of visual stimulus and fluorescent response features, sliding time/frame windows, and dimensionality reduction to generate inputs in a model to classify multiple neuronal cell types in a single microscopy recording. We originally hypothesized that the K-nearest neighbors and/or neural network implementations would produce the strongest classification performance due to the algorithms’ ability to flexibly fit nonlinear feature spaces. Due to the imbalanced nature of the dataset, with five classes total and one class making up nearly 50% of the data, balanced accuracy is a better indicator of model performance than accuracy. Classifying cells via random chance would yield a balanced accuracy of 20%. Our best cell type classifier, a convolutional neural network optimized for time series classification, gives us an accuracy score of 70.6% and balanced accuracy of 53.7%.