UC San Diego

The field of subcellular spatial transcriptomics has allowed researchers to better understand the spatial organization of transcripts in various cell types. However, attempting spatial transcriptomic experiments on neurons has proven difficult due to the complex interconnected network that clutters their boundaries and complicates the mapping of transcripts to their parent neuron. To address this problem, I made advancements on three fronts: clarification methods, advanced imaging, and image processing algorithms. By attempting methods of sparse neuron culturing, combinations of immunofluorescent protein stains, and several microscopy techniques, I created images with minimal intersecting neurons for easier segmentation. Furthermore, I combined methods such as machine learning, graph theory, and novel algorithms to develop a python package, Mosaic, to computationally differentiate individual neurons in the microscopy images. Mosaic includes image preprocessing steps to prepare the image for segmentation. An optional gap bridging algorithm is included to help bridge any gaps caused by punctated immunofluorescence staining that break the continuous illumination of the neuron. The core of the software package consists of both a soma segmentation algorithm built upon Stardist and CellPose, two state-of-the-art cell instance segmentation packages, and a reconstruction algorithm that returns the best instance trace of every neuron. The tool produces accurate segmentation masks from which further spatial transcriptomics data analysis pipelines can be mapped to. In conjunction with my colleagues' python toolkit for the analysis of subcellular spatial transcriptomics data, Bento, we will enable the analysis of subcellular RNA localization in neurons in fully connected networks of neurons.