Conférence de Gabriel Leclerc, étudiant à la maîtrise en génie informatique

High-throughput analysis of calcium signaling in neurons is a key element for the understanding of specific functions of neuronal sub-compartments. Fluorescent calcium sensors and optical microscopy enable the recording of time series of calcium (Ca2+) dynamics in living neurons with high spatiotemporal resolution. We propose to develop an approach based on deep learning to automatically detect and segment miniature spontaneous calcium transients (mSCTs) in fluorescence microscopy videos of living cultured hippocampal neurons.

Considering the very small proportion of positive events that are detected in Ca2+-imaging data, we are following a positive unlabeled (PU) learning schema. This framework allows inferring models to detect and segment mSCTs on dendritic shafts and spines (foreground) from a limited set of positive mSCTs labels (P) augmented with a variable set of unlabeled instances (U) belonging to the foreground of the Ca2+ imaging data. This improves the detection performance using fewer labels and takes advantage of events sparsity found in these data. Varying the proportions of P and U instances, we trained a UNet neural network that outperforms a threshold-based image processing algorithm that was used to generate the training targets.

This approach not only provides the possibility to perform high-throughput analysis of mSCTs but also enables the detection and segmentation of low-intensity events that would have been otherwise undetected.