Electron Microscopy Denoising

Removing directional noise — scan lines, streaks, reconstruction artifacts — from EM volumes with the routed pipeline.

The Challenge

EM volumes typically mix:

• Shot/read noise — random, pixel-independent
• Directional structure — scan lines from the acquisition, detector streaks, and in tomograms the directional artifacts of the reconstruction

Plain Noise2Void removes the random component but can leave the directional one: neighboring pixels share correlated noise, and the blind-spot network copies it from context. Whether that matters for your volume is exactly what ASN2V measures before training.

Dense EM (FIB-SEM / COSEM-style)

Dense cellular material with bright resin background — bg_side="light":

from autoStructN2V.pipeline import run_pipeline

summary = run_pipeline({
    "input_data": "cosem_stack.tif",
    "output_dir": "./results",
    "experiment_name": "cosem_routed",
    "mask": {"source": "extractor",
             "extractor": {"bg_side": "light"}},
})
print(summary["branch"], "|", summary["route_reason"])

Cryo-ET Tomograms

On the real tomograms tested in the paper, flatness-only background selection worked best — bg_side="off":

"mask": {"source": "extractor", "extractor": {"bg_side": "off"}}

Tomogram noise is often anisotropic (missing wedge, reconstruction direction), which is precisely the directional correlation the spine mask captures.

What to Expect

• Seconds after start you have the routing decision and (on the StructN2V route) the discovered mask — inspect both before the training run commits GPU time.
• On the PhantEM EM benchmark, the routed method reaches a five-seed mean Pearson of 0.668 vs 0.582 for plain N2V, matching an oracle mask built from the true noise under identical design rules.
• On volumes without usable directional structure, the router falls back to plain N2V — the correct treatment, discovered automatically.

Always look at the residual (create_difference_map(noisy, denoised)): it should contain noise, not biology.

Try It on Open Data

The full pipeline runs out of the box on the PhantEM benchmark pairs (six EM volumes with clean ground truth, bg_side="off") — the Basic Usage tutorial walks through it step by step. The paper's split-sum tilt-series experiment uses the open Zenodo record 14922032 (Kim et al. 2025).

Adapting to Your Data

• Set bg_side for your contrast: 'light' (bright resin), 'dark' (dark background), 'off' (flatness-only)
• If the mask looks implausible, check bg_side first, then point bg_box at a known-empty region
• See Configuration for all options — and prefer the presets/defaults; they are validated operating points