ALTEA

Autonomous Learning for Tomographic Ensembles and Attributes

A reproducible, provenance-tracked Python pipeline for automated quality control, segmentation and 3-D morphometry of tomographic image stacks such as FIB-SEM. Built with porous and heterogeneous materials in mind (MOFs, battery electrodes, catalysts, membranes, rock), but general to any two-phase or multi-phase volumetric acquisition.

---

Why

Analysing FIB-SEM tomography today typically involves two costly, poorly reproducible steps: an operator manually discards a large fraction of acquired slices (e.g. 500–600 acquired, 300–400 kept) by visual inspection, and the image treatment is tuned by eye per sample "like Photoshop", with no single procedure that applies across samples. ALTEA turns both into deterministic, config-driven, versioned operations, and records exactly what was done to every volume.

ALTEA is a reproducible orchestration layer, not a single model. Every stage is explicit, every parameter lives in a config file, and every run emits a provenance record (software and dependency versions, input data hash, per-stage parameters and metrics). The same input always yields the same result — and you can prove it.

Pipeline

load → quality control → drift correction → preprocessing → segmentation → morphometry

Stage	What it does
io	Load FIB-SEM stacks (multi-page TIFF or slice directories) with voxel spacing
qc	Per-slice sharpness / brightness / drift / curtaining scores; reproducible, threshold-driven slice selection (replaces manual curation)
align	Rigid drift correction by phase cross-correlation
preprocess	Deterministic denoising and contrast normalization — every parameter recorded
segment	Pluggable backends: classical (Otsu, watershed) and learned (random-forest pixel classifier; optional deep U-Net)
morphometry	Porosity, specific surface area, pore-size distribution, connectivity/percolation, geometric tortuosity — all voxel-size-aware
acquire	Convergence analysis: how few slices suffice for a target accuracy (basis for cost-aware acquisition)

Documentation: https://altea.readthedocs.io

Install

pip install altea            # core
pip install "altea[deep]"    # + optional deep-learning segmentation backends
pip install "altea[dev]"     # + test / lint tooling

Quickstart

No data required — run the synthetic demo:

altea demo --output runs/demo

Or on your own stack:

altea run --input stack.tif --config configs/fibsem_default.yaml --output runs/sample1

In Python:

from altea import Pipeline
from altea.datasets import make_porous_volume, add_acquisition_artifacts

clean, ground_truth = make_porous_volume(porosity=0.35)
raw = add_acquisition_artifacts(clean, blur_slices=(8,), charge_slices=(15,))

results = Pipeline.from_yaml("configs/fibsem_default.yaml").run(raw, output_dir="runs/demo")

m = results["morphometry"]
print(m.porosity, m.specific_surface_area, m.tortuosity)
print(results["qc_report"].summary())      # which slices were dropped, and why

Segmentation backends

Backends share one interface and self-register, so switching is a one-line config change:

segment:
  backend: pixel_rf      # otsu | watershed | pixel_rf | unet
  params: {sigmas: [1, 2, 4], n_estimators: 200}

Add your own by subclassing SegmentationBackend and decorating with @register_backend. Nothing else in the pipeline changes.

Reproducibility

Every run writes provenance.json:

{
  "altea_version": "0.1.0",
  "dependency_versions": {"numpy": "...", "scikit-image": "..."},
  "input_hash": "sha256:...",
  "config": {...},
  "stages": [{"name": "qc", "params": {...}, "metrics": {...}}, ...]
}

Roadmap

• Cost-aware autonomous acquisition. The acquire module already quantifies the accuracy-vs-slice-count trade-off. The next step is an active policy that decides on the fly how many sections to acquire to reach a target uncertainty, reducing destructive beam time.
• Deep segmentation backends (2-D U-Net / nnU-Net wrapper) under altea[deep].
• Sparse-view reconstruction of discarded slices.

Status

Alpha. Interfaces may change before 1.0.

Citation

If you use ALTEA in your research, please cite the archived software:

Bravo-Abad, J. (2026). ALTEA: Autonomous Learning for Tomographic Ensembles and Attributes (v0.1.0). Zenodo. https://doi.org/10.5281/zenodo.21442516

@software{bravoabad_altea_2026,
  author    = {Bravo-Abad, Jorge},
  title     = {{ALTEA: Autonomous Learning for Tomographic Ensembles
               and Attributes}},
  year      = {2026},
  version   = {0.1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21442516},
  url       = {https://doi.org/10.5281/zenodo.21442516}
}

Machine-readable metadata is in CITATION.cff.

License

MIT — see LICENSE.