SiDoLa-NS

SiDoLa-NS — "Simulate, Don't Label – Nervous System" — is a suite of instance segmentation networks for nervous-system tissue from William Buchser's group in the Department of Genetics and the McDonnell Genome Institute at Washington University in St. Louis.

Spatial omics platforms map transcripts and proteins at subcellular resolution, but attributing those molecules to cells and to named anatomical regions requires segmenting the accompanying image. In nervous system tissue that is unusually hard: nerve cross sections are dense and irregular, and delineating spinal laminae or brain atlas regions has traditionally meant manual registration. Supervised networks could do it, but hand annotation at that scale is slow, inconsistent, and imposes annotator bias on the boundaries the model learns.

SiDoLa-NS removes the annotation step entirely. Rather than labeling real micrographs, the authors build a biophysical simulator in Blender that constructs tissue from parameterized 3D geometry — cells, then regions of cells, then whole sections — and renders it through a virtual microscope with depth of field, noise, tears, warping and skew. Because the geometry is known before rendering, ground-truth polygons come out of the simulator exactly. This is the opposite approach to segmentation models such as Cellpose, trained on large collections of real, human-annotated images.

#Key Features

• Trained only on simulation: Both training and hyperparameter optimization use synthetic images, so no memorization or batch-effect leakage from the test tissue is possible.
• Three spatial scales in one suite: Micro models segment nuclei, cells and axons; the meso model finds cell clusters, laminae and nerve fascicles; macro models assign named atlas regions across a whole section. The scales are fused at inference so each cell carries a region label.
• Rare and difficult cases are over-represented: The simulator deliberately generates out-of-focus, broken, warped and extremely high-density tissue, the conditions where segmentation networks trained on curated real images fail.
• Perfect, updatable ground truth: Object boundaries are defined by the geometry before optical aberrations are applied, and the labeling schema can be changed and the whole dataset regenerated — impractical when re-labeling 100,000 images by hand.

#Technical Details

Five models are released. Micro-CNS and Micro-PNS segment nuclei and axons in the central and peripheral nervous system; Meso-pSN detects fascicles in porcine sciatic nerve; Macro-mCB and Macro-mSC map atlas regions in mouse coronal brain and spinal cord. The coronal brain checkpoint is filed on GitLab as SiDoLa-NS_Macro-mBR.pt. All are built on the Ultralytics YOLO framework — v9c and v9e for the micro models, v8n-seg and v8l-seg for the macro models — trained for 10 to 200 epochs at image sizes between 448 and 640 pixels, with neural architecture search selecting the settings. Training sets are generated at about 4,060 images per hour: 4,353 image pairs and 1.96 × 10⁶ objects for the first micro-scale brain set, 144,438 training images for the macro brain set, and 7,198 covering 2.78 × 10⁵ objects for the sciatic nerve set. On simulated validation data, mAP50 is 0.715 for Micro-CNS and 0.977 for Macro-mCB, Macro-mSC reaches a top precision of 0.96 at 0.92 confidence, and F1 exceeds 0.9 for Micro-PNS and reaches 1.0 for Meso-pSN. Macro-mCB holds mAP50 above 0.783 with up to 70 percent noise added; Macro-mSC stays above 0.60 out to 40 degrees of section rotation.

#Applications

The suite sits in front of a spatial omics analysis, converting an H&E, Nissl, immunofluorescence or DAPI section into cells with region assignments so that transcripts can be aggregated by anatomy rather than by unsupervised expression clusters. The paper uses it to rank region-specific genes on 10X Visium HD mouse brain and Xenium spinal cord data, to compare cell counts and region confidence between wild-type and CRND8 Alzheimer's model mice, and to count axons and nuclei per fascicle in a porcine sciatic nerve. Micro-CNS also transferred to a DAPI-stained lung adenocarcinoma image with no parameters changed from the brain settings, so the nuclei model is not confined to neural tissue. Code, notebooks and four of the five checkpoints ship on GitLab under AGPL-3.0; the simulated training datasets are on HuggingFace under Apache-2.0, and two of those also carry model artifacts — Meso-pSN holds the fifth checkpoint, meso_model.pt, alongside an ONNX export, and Macro-mCB an ONNX export. A SiDoLa PyPI package runs inference on a folder of images.

#Impact

SiDoLa-NS argues concretely that for segmentation of structured tissue a well-parameterized forward simulator can substitute for an annotated corpus — and can exceed it, because the simulator supplies exact boundaries and unlimited rare cases that human annotators cannot. Its top-down design also sidesteps the double-dipping problem in spatial omics, where regions defined from molecular markers are then tested for differential expression of those same markers. The authors identify three limitations: the user must specify the scale at which each model is applied, the macro models are trained on a single reference atlas section rather than the full volume, and instance segmentation is an imperfect fit for atlas regions, which are non-overlapping by definition. The work remains a preprint, posted under CC BY-NC-ND.