Section: New Results

Convolutional Neural Networks algorithms for calcium signal segmentation in astrocytes in 3D lattice light sheet microscopy

Participants : Anais Badoual, Charles Kervrann.

Astrocytes, glial cells of the central nervous system, are detectors and regulators of neuronal information processing. It is established that neuronal synapses are physical sites of intercellular contact that transmit and transform information in a very rapid and flexible way, playing a pivotal role for learning and memory formation as well as neurological diseases of the mammalian brain. Astrocytes are thought to integrate neuronal inputs and modulate information transfer between neurons. In particular, cytoplasmic calcium signaling in astrocytes is believed to be crucial for astrocyte-neuron communication. However, quantification of intracellular calcium signals in astrocytes is hindered by the complexity of their cell shape, that consists of a cell body sprouting a highly ramified set of large to very fine protrusions called processes. Until recently, the quantification of intracellular propagation of calcium signal in astrocytes with fluorescent calcium indicators has been restricted to two dimensions, either 2D cell cultures or 2D slicing of a 3D setup. However it is not clear what amount of information is lost by ignoring the 3rd dimension in these experiments. The emergent 3D Lattice Light Sheet Microscopy (LLSM) is a powerful and promising technology (voxel size: 250nm x 250nm x 700nm; acquisition time: 200 frames per second) to give a much more complete and refined view of the dynamic behavior of calcium signaling in astrocytes inside living brain slices and in the intact mouse brain in vivo. Unfortunately, we lack image analysis tools to locate, segment, track and quantify the propagation of those 3D calcium signals in very ramified cell shapes.

In this context, we have started to develop an image processing tool for neurobiologists that 1) detects and segments calcium signals in 3D+time LLSM images, and 2) classifies these signals based on their 3D space-time morphological characterization. To do so, we focus on 3D convolutional network and machine learning techniques.

Collaborators: V. Nägerl and M. Arizono (Interdisciplinary Institute for Neuroscience, Bordeaux),

                          H. Berry and A. Denisot (EPC beagle , Inria Rhone-Alpes).