Person:

Park, Core Francisco

Recent advances allow sustained rapid 3D functional imaging of large numbers of neurons inside behaving animals. To decode neural activity, individual neurons must be segmented and tracked. This can be challenging inside brains that rapidly move and deform within flexible, behaving animals. No general method is able to effectively solve this problem. We present an analysis method based on a convolutional neural network (CNN) with specific enhancements, which we apply to freely moving C. elegans. For a traditional CNN to track neurons across brain images with different postures, the CNN must be trained with ground truth (GT) annotations of similar postures. When these postures are diverse, adequate numbers of manual GT annotations can become prohibitively large to generate. We introduce `targeted augmentation', a method to automatically synthesize reliable annotations from a few manual annotations. Our method effectively learns the internal deformations of the brain. These learned deformations are used to synthesize annotations for new postures by deforming the manual GT annotations of similar postures. When synthetic annotations are added to training datasets, the need for manual GT annotation and final proofreading are reduced. We provide a GUI for our method in an end-to-end pipeline from manual GT annotation to final proofreading. We demonstrate our method by segmenting and tracking neurons -- where individual neurons can be represented either as points or 3D volumes -- in a variety of C. elegans strains across multiple animals. In a study of olfactory processing, our method uncovered rich patterns in interneuron dynamics in freely moving animals exposed to periodic odor pulses, including dynamic on-off switching of neuronal entrainment to time-varying stimuli.

Connectomics provides nanometer-resolution, synapse-level maps of neural circuits to understand brain activity and behavior. However, few researchers have access to the high-throughput electron microscopes necessary to generate enough data for whole brain or even whole circuit reconstruction. To date, machine-learning methods have been used after the collection of images by electron microscopy (EM) to accelerate and improve neuronal segmentation, synapse reconstruction and other data analysis. With the computational improvements in processing EM images, acquiring EM images has now become the rate-limiting step in automated connectomics. Here, in order to speed up EM imaging, we integrate machine-learning into real-time image acquisition in a single-beam scanning electron microscope. This SmartEM approach allows an electron microscope to perform intelligent, data-aware imaging of specimens. SmartEM saves time by allocating the proper imaging time for each region of interest – first scanning all pixels rapidly, then rescanning more slowly only the small subareas where a higher quality signal is required. We demonstrate that SmartEM achieves up to a $\sim$7-fold acceleration of image acquisition time for connectomic samples using a commercial single-beam SEM in samples from nematodes, mice and human brain. We apply this fast imaging method to reconstruct a portion of mouse cerebral cortex with an accuracy comparable to traditional electron microscopy.