Unknotting Neurons: Automating the Segmentation of Mouse Hippocampal Culture Imagery using Spatial Transcriptome Profiling

Abstract

Recent developments in imaging of RNA, including multiplexed, error-robust fluorescence \textit{in situ} hybridization (MERFISH) have allowed for the mapping of the subcellular distributions of hundreds to thousands of transcribed genes at once. The study of spatial transcriptomics is particularly interesting in the context of neurons, where the need for translation in neurites leads to mRNAs localizing far from their site of transcription. Automated cell segmentation in neurons presents a challenge for traditional image segmentation techniques due to irregular cell shapes and intersecting neurites. Here, we present a program for automated cell segmentation in mouse hippocampal culture imagery informed by the use of MERFISH data. This program makes use of traditional image processing techniques and machine learning methods in order to classify and segment images, producing highly accurate segmentation of somata and dendrites. We introduce embedding schemes to help distinguish cells and compare the localization of mRNA within processes and the distribution of RNA along dendrites is modeled, with attention to differences in decay rates between species. We discuss how these embeddings and models may contribute to future neuron segmentation procedures.