Biophysical Simulation Enables Multi-Scale Segmentation and Atlas Mapping for Top-Down Spatial Omics of the Nervous System

Abstract

Spatial omics (SO) has produced high-definition mapping of subcellular molecules (like transcripts or proteins) within tissue samples. Mapping transcripts to anatomical regions requires segmentation, but even segmenting nuclei remains challenging for tissues like nerve cross sections, let alone for larger regions such as within the spinal cord. Neural networks could address this but need extensive human annotations--a bottleneck. We present SiDoLa-NS (Simulate, Dont Label - Nervous System), an image-driven (top-down) approach to SO analysis in the nervous system. We utilize biophysical properties of tissue architectures to design synthetic images mimicking tissue samples. With these in silico samples, we train supervised instance segmentation convolutional neural networks (CNNs) for nucleus segmentation, achieving precision and F1-scores > 0.95. We take this a step further with generalizable CNNs that can identify macroscopic tissue structures in the mouse brain (mAP50 = 0.869), spinal cord (mAP50 = 0.96), and pig sciatic nerve (mAP50 = 0.995). Short SummaryThe SiDoLa-NS micro-, meso-, and macro-scale models are generalizable, supervised CNNs for neuronal segmentation in cell to tissue-level contexts. SiDoLa-NS is novel in its combination of three core ideas: it is top-down (image first), trained solely on synthetic images, and it is multi-scale. The tool is validated on brain, spinal cord, and sciatic nerve for advanced segmentation tasks.