Reconstruction and Analysis of Mitochondrial Morphology and Distribution in Neocortical Neurons using 3D Deep Learning

Abstract

Mitochondria play a crucial role in the functioning of neurons by synthesizing adenosine triphos- phate (ATP), a compound used to deliver energy intracellularly, and by buffering intracellular Ca2+. Both are necessary for neuron-to-neuron signal transmission via synapses. Mitochondria are highly dynamic organelles, constantly changing through fusion and fission events, moving intracellularly and often forming long filaments in order to adapt to a changing cell environment. The morphology and location of mitochondria can give insight into the amount of energy required for processes to proceed in different parts of the cell. Previous studies, which have focused on small fragments of the hippocampus, have identified the role of mitochondria in the creation of new synapses and responding to hypoxic conditions. Differences have also been observed in the distribution and level of connectivity of mitochondria between dendrites, axons and cell bodies. Although the presence of mitochondria in dendritic spines has been disputed, it has been observed in CA3. Studies of larger areas of various brain regions are necessary to thoroughly describe the distribution of neural mitochondria, in order to gain more insight into the functioning of this crucial organelle. Many studies thus far have observed single neurons in vivo or depended on manual reconstructions of electron microscopy (EM) images. In the recent years, various semiautomatic and fully automatic approaches have been explored to identify mitochondria in cells. The deep learning approach to semantic segmentation of mitochondria in EM images developed by Dorkenwald et al. (2017) has achieved highest accuracy. We apply a similar deep learning approach to identify mitochondria in a 196 x 130 x 40 μm3 TEM dataset of mouse primary visual cortex (V1) collected by the Allen Institute as part of the iARPA MICrONS program. The dataset contains cell reconstructions and single-cell functional data, which can be used with our mitochondrial profiles. We use a 3D neural network to classify each dataset voxel as either mitochondrion or background and reconstruct individual mitochondria objects using connected components. Applying the efficient CPU implementation of the RSUNet we achieve the F1 score of 0.867, with the inference speed is 3.5 CPU hours per gigavoxel. We reconstruct over 900,000 mitochondria objects in the volume. We assign each mitochondrion to a cell segmentation and morphological cell part (axon, dendrite, soma) in the volume. For cells with somas in the volume we calculate the distance from soma center of each mitochondrion object and calculate the mitochondrion volume density distribution as function of distance from soma center. We consider mitochondria volume and length variability, as well as the proportion of axon and dendrite lengths containing mitochondria across cell types. We observe filaments spanning the entire dataset (over 170μm in length) in apical dendrites, exceeding the length of the largest filaments reported so far by more than a factor of three. Surprisingly, we also find small mitochondria in dendritic spines in V1.