Biophysical Simulation Enables Multi-Scale Segmentation and Atlas Mapping for Top-Down Spatial Omics of the Nervous System
Spatial omics (SO) has produced high-definition mappings of subcellular molecules (like transcripts or proteins) within various tissue samples. Most applications of SO are molecule-driven i.e. the spatial distributions of transcripts are used to make distinctions between samples. However, such inferences do not automatically utilize brain atlas regions. Here, we present SiDoLa-NS (Simulate, Dont Label - Nervous System), an image-driven (top-down) approach to SO analysis in the nervous system. We utilize the biophysical properties of tissue architectures to design synthetic images mimicking tissue samples. With these in silico samples, we train supervised instance segmentation models for nucleus segmentation, achieving near perfect precision and F1-scores > 0.95. We take this a step further with generalizable models that can identify macroscopic tissue structures in the mouse brain (mAP50 = 0.869) and spinal cord (mAP50 = 0.96) and pig sciatic nerve (mAP50 = 0.995). SiDoLa-NS provides a framework in applying and analyzing SO data that leverages high-definition images that are taken alongside spatial omics pipelines. Notably, SiDoLa-NS provides solutions for common challenges in supervised model training, including but not limited to annotator bias, limited generalizability, and production of massive, high-quality training sets.

Short SummaryThe SiDoLa-NS micro-, meso-, and macro-scale models are generalizable, supervised models 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.