In this dissertation, methods for characterizing cells based on their mechanical phenotypes are described. A novel microfluidic channel design is presented and data are gathered as cells pass through undulations in the channel. Deep learning methods are applied to the data in a new approach for classifying cells solely based on their mechanical properties. First, in a supervised deep learning approach, a highly interpretable random forest was created and trained to extract the most influential features for cell classification. Feature attributions of the random forest were uncovered using Shapley values. Analysis of the most influential features revealed by the Shapley values highlighted the importance of temporal features, such as change in aspect ratio over time, in classifying cells. This led to the development of a powerful convolutional recurrent neural network, which dramatically improved classification accuracy to over 90% when using five-fold cross-validation.
An unsupervised deep learning approach for cell classification was then explored for problems where classes of cells are unknown a priori. Unsupervised classification was first tested using manually extracted features and traditional clustering algorithms. However, performance was significantly improved with the development of a variational autoencoder (VAE), which extracted higher dimensional features. Additionally, the encoder for the VAE was made into a classifier using a clustering loss function. This trained network exhibited an unsupervised clustering accuracy of up to 80% when thresholding the top ~10% of predictions.