No abstract

No abstract

No abstract

Reconstruction of serial section electron microscopy (ssEM) data greatly facilitates neuroscience research, but such reconstruction is computationally expensive. Informative data about physiological functions can in theory be obtained from ssEM datasets by extracting distinct cellular structures without large-scale reconstruction, but an efficient method is needed to accomplish this. Here, we developed a Region-CNN (R-CNN) based deep learning method to identify, segment, and reconstruct synapses and mitochondria from ssEM data. We applied this method to explore the changes in synaptic and mitochondrial configuration in the auditory cortex of mice subjected to auditory fear conditioning. Upon reconstructing over 135,000 mitochondria and 160,000 synapses, we found that fear conditioning significantly increases the number—while decreasing the size—of mitochondria, and also noted that it promoted the formation of multi-contact synapses comprising a single axonal bouton and multiple postsynaptic sites from different dendrites. Combinatorial modeling indicated that such multi-dendritic synapses increased information storage capacity of new synapses by over 50%, representing a synaptic memory engram. Our method achieved high accuracy and speed in synapse and mitochondrion extraction, and its application revealed structural and functional insights about cellular engrams associated with fear conditioning.