PDF, English Download (41MB) | Terms of use

Abstract

Detecting repeating firing motifs of neuron groups (so-called neuronal assemblies) and cell segmentation in calcium imaging, a microscopy technique enabling the observation of neuronal activity, are two fundamental and challenging tasks in neurophysiological data analysis. In this thesis, three novel approaches are presented, which use machine learning to tackle both problems from different perspectives. First, SCC is presented for the detection of motifs in neuronal spike matrices, which are gained from calcium imaging data by cell segmentation. SCC uses sparse convolutional coding and outperforms established motif detection methods by leveraging sparsity constraints specifically designed for this data type combined with a method to avoid false-positive detections. Second, LeMoNADe is the first method ever to detect spatio-temporal motifs directly in calcium imaging videos, eliminating the cumbersome extraction of individual cells. It is a variational autoencoder framework tailored for the extraction of neuronal assemblies from videos and matches the performance of state-of-the-art detection methods requiring cell extraction. Although LeMoNADe enables the detection of neuronal assemblies without previous cell extraction, this step is still essential for a wide range of downstream analyses. Therefore, the third method, DISCo, combines a deep learning model with an instance segmentation algorithm to address this task from a new perspective and thereby outperforms similarly trained existing models.