NUScope-TOMO

Open FIB-SEM slice-and-view tomography — from raw slices to quantified, chemically resolved 3D microstructure.

Created by Roberto dos Reis (@rmsreis) · NUANCE Center, Northwestern University · NUANCE-IT/NUScope-TOMO

NUScope-TOMO is an open alternative to Thermo Fisher Avizo for FIB-SEM in materials science. It covers import, pre-processing (with a guided FIB-SEM wizard), AI segmentation, quantification, triple-phase boundaries, pore/particle networks, transport simulation, visualization/animation and recipes. It also adds correlative EDS chemistry. Everything is available as code, as reproducible YAML recipes and in a browser-based UI. PLAN.md maps each Avizo capability to its module.

FIB-SEM sample 335 (812 slices, 20 × 5 × 5 nm voxels) processed by examples/fib_sem_335.yaml in under 5 minutes: dark phase with a grey-value cut-away (left) and 1 414 separated particles coloured by equivalent diameter (right).

Materials analyses: steady-state diffusion through the matrix phase (left, tortuosity factor 1.4 along Z, 3.8–3.9 across) and the contact network of the dark-phase particles (right, mean coordination 3.0).

Install

pip install nuscope-tomo            # or "nuscope-tomo[ai]" for Noise2Void / U-Net (PyTorch)

For development, in a clone: uv pip install -e ".[dev]". With NUScope installed, the same commands are also available as nuscope tomo .... With NUANCE-MCP installed, NUScope-TOMO adds tools for tomography projects (tomo_project, tomo_table, tomo_volume, tomo_phase_fractions, tomo_ops, tomo_run, tomo_job).

Use

nuscope-tomo run examples/fib_sem_335.yaml          # full pipeline on a slice-and-view stack
nuscope-tomo studio projects/fib_sem_335            # open http://127.0.0.1:8765
nuscope-tomo ops                                    # list operations and parameters
nuscope-tomo watch /acq/folder -p projects/run --voxel 20 5 5 --model rf.pkl   # live mode
python examples/beyond_avizo.py                # data health, Noise2Void, U-Net, RVE
nuscope-tomo compare A=projects/a/run B=projects/b/run --out cohort
python examples/chemical_fusion.py             # sparse DKL vs pan-sharpening benchmark
python examples/self_driving.py agent          # autonomous run on the digital twin
python examples/showcase.py                    # figures, renders and animation

From Python:

from nuscope_tomo import Volume
from nuscope_tomo.preprocess import auto_crop, align, destripe, denoise_nlm
from nuscope_tomo.segment import random_forest
from nuscope_tomo.analyze import measure_objects

vol = Volume.from_zarr("raw.zarr")
vol = denoise_nlm(destripe(align(auto_crop(vol))["volume"]))
phases = random_forest(vol, scribbles="annotations.json", classes=3)["volume"]

Remote use: run nuscope-tomo studio --host 0.0.0.0 on the workstation that holds the data and open it from any laptop. The browser only receives slices and a downsampled 3D preview.

Design notes

• Out-of-core first. Volumes are Dask arrays over Zarr; slice-wise operations stream. Connected components, watershed and surface meshing run in memory, on a binned volume or an ROI for very large data.
• Anisotropy everywhere. FIB-SEM voxels are typically 5–10 nm in XY and 10–50 nm in Z. Distances, filters, surface areas and renders all use physical spacing.
• Validated against ground truth. pytest runs 26 tests on synthetic phantoms. They cover drift recovery, stripe removal, sphere volume, area and sphericity (within 1–2 %), percolation and tortuosity, random-forest accuracy, watershed splitting, TPB length on a straight triple line, transport through a straight channel (τ = 1, D_eff = ε), the neck of two fused spheres, slice-thickness detection, integral-range error prediction, Noise2Void, U-Net, live ingestion, the advisor, cohort statistics, sparse DKL, pan-sharpening, the AutoScript dry-run and safety guards, the agent on a twin, EDS registration and phase recovery, the recipe engine and the Studio API.
• Honest alignment. Correlation cannot separate stage drift from microstructure that is inclined to the milling axis, which is especially a problem with thick slices. align therefore detects stacks that are already registered and leaves them alone, and supports fiducial ROIs and a high-pass drift model.

Project layout

nuscope_tomo/
  core/        Volume, operation registry, recipes + provenance
  io/          slice import/export
  preprocess/  geometry, alignment, artifact correction, denoising
  segment/     thresholds, random-forest classifier, object separation
  analyze/     object morphometry, phase metrics
  chem/        EDS chemical tomography (+ simulator)
  viz/         GPU rendering, animation, figures
  studio/      FastAPI server + WebGL2 front-end
  cli.py       `nuscope` command
examples/      recipe for the sample dataset, showcase script
tests/         ground-truth tests

See PLAN.md for the Avizo feature map and roadmap: U-Net segmentation, Noise2Void, 3D multi-volume registration, permeability, native .h5oina / .bcf EDS readers.

Licence

MIT, see LICENSE. Copyright (c) 2026 Roberto dos Reis (@rmsreis).

Citation

If NUScope-TOMO helps your research, please cite it using CITATION.cff. Author: Roberto dos Reis (@rmsreis).