Neuronal Circuit and Synapse Analysis with Deep Neural Networks. In Data Science & Engineering Master of Advanced Study (DSE MAS) Capstone Projects

Biomedical image segmentation has led to numerous breakthroughs in neuroscience research by helping scientists map the brain. Brain mapping could be the key to understanding the progression of neurodegenerative diseases. Currently, the most accurate neural segmentations and synapse detections are annotated manually. However, manual annotation of synapses is unsustainable given the very large size of electron microscopy datasets of the brain. Deep learning tools such as CDeep3M can produce automated synapse segmentations on serial block-face scanning electron microscopy (SBEM), but the segmentations are relatively inaccurate. This is due to the numerous barriers including the inherent heterogeneity of neurons and their synapses and the limited tools for assessing and analyzing ultrastructure in molecularly defined synapses. Our developed solution for improving the automated segmentations of synapses, known as CDense3M, could help neuroscientists discover new insights in neurodegenerative diseases such as Alzheimer’s and Parkinson’s disease. We implement and apply two different modeling components to perform 3D segmentations and labeling to improve the accuracy of synaptic density segmentations. The first modeling component creates different algorithms for 3D concatenation, labeling, and filtering. As for 3D filtering, we also implement multiple models that use erosion and determine the statistical thresholds for 3D vesicle filtering on 3D objects. In each step of our 3D modeling, the new data is based off of previous modeling and as a result the data in each step of modeling becomes the new data for the next step. The evaluation for our image processing including 3D labeling and 2D and 3D vesicle filtering models is through the visualization and using the biological behavior of known neuronal structures. This improves the accuracy of defining the presynaptic areas based on the vesicle functions. The second modeling component applies flood-filling networks, a class of 3D convolutional neural networks, to perform cell-body segmentation on a 3D image volume. We train a custom flood-filling network model on membrane segmentations produced by the CDeep3M automated segmentation tool. The membrane model is more generalizable at segmenting cell bodies than the standard FIB-25 model trained on raw electron microscopy data from fruit flies. The improvement in the accuracy of presynaptic area recognition from applying both modeling components results in a 10% increase in precision and recall for the voxel-wise classification of synapses in a 3D image volume of the Nucleus accumbens from lab mice. This capstone project was conducted to fulfill the requirements for the Master of Advanced Study Degree in Data Science and Engineering at the University of California, San Diego.