Geometry Import and Segmentation

Geometry Import & Segmentation

The pipeline starts from nerve geometry. golgi accepts ready-made surfaces, pre-built bundles, and — for raw imaging — provides promptable segmentation and 3-D reconstruction to turn microscopy / µCT into a meshable nerve.

Source: golgi/segmentation/ (image.py, segmenter.py, reconstruct3d.py, deform.py), golgi/projects/uct_route.py (upload), and Study.import_nerve for surfaces.

---

Ways to get geometry in

Path	Input	Notes
Surface import	STL / NAS / OBJ	a watertight nerve surface; the fastest route. Study.import_nerve(...) or the Import wizard's STL surface tile.
µCT / histology bundle	a golgi bundle (epineurium + per-fascicle endoneurium)	produced by segmentation; carries multiple fascicles.
Segment from images	TIFF / DICOM / NIfTI / NRRD / PNG / …	segment slices in-app, then reconstruct to surfaces.

On surface import, golgi computes the global PCA frame, per-triangle surface quality, and topology stats (component count, watertightness, bounding box). scale_factor converts file units to metres (default mm→m). golgi assumes a watertight surface — repair broken meshes upstream (MeshLab / Blender); see Meshing.

---

Promptable segmentation

The Segment µCT slice dialog segments image slices interactively:

• Image loading (image.py) reads many formats — TIFF (lazy per-slice, multi-GB-safe), DICOM (single file or series), NIfTI, NRRD, MetaImage, Analyze, JPEG2000, and ordinary PNG/JPEG — capturing voxel size where available.
• Segmenter (segmenter.py) uses a MedSAM2 / SAM-style backend when a model checkpoint is available: propose-all ("everything") masks, or segment-at refinement from positive/negative point clicks and a box. Each proposal carries a confidence score, bounding box, and area.
• Without a model checkpoint, golgi falls back to a stub segmenter so the workflow still runs; the checkpoint is an optional install (see Installation).

Accept and label masks per slice (e.g. fascicle IDs), and golgi assembles them into a nerve.

3-D reconstruction

reconstruct3d.py turns segmented slices into surfaces (mm units):

• extrude_single_slice — one annotated cross-section + a thickness → a prismatic nerve (useful for extruded-cross-section studies and as a control against full 3-D reconstruction).
• reconstruct_stack — a slice range (gaps filled) → per-fascicle surfaces via 3-D connected components + marching cubes, with optional smoothing.

The result is the multifascicular geometry (epineurium + per-fascicle endoneurium) the rest of the pipeline meshes and solves.

Uploads

Large images upload through a streaming route (uct_route.py) — multipart POST written to the project in 64 KB chunks (so multi-GB DICOM series don't blow up memory), with progress reported back to the GUI.

---

See also

GUI Walkthrough · Meshing · Fiber Populations & Trajectories · Getting Started