PhantEM

Realism-validated synthetic benchmark pairs for denoising volumetric electron microscopy.

From a real acquisition to a validated benchmark pair: phantom rendered from the volume's own segmentation, noise fitted to its own background

The Problem

Denoising methods for volumetric EM are routinely evaluated on synthetic data whose resemblance to real data is asserted rather than measured. Real cryo-ET and FIB-SEM volumes have no clean counterpart to score against — and their noise is spatially correlated, anisotropic, and correlated between slices, none of which white noise reproduces. A benchmark built on unrealistic noise can invert the ranking of the very methods it is supposed to compare.

The Solution

PhantEM generates benchmark pairs in which both halves are anchored to a real acquisition:

• The clean volume is a phantom rendered from a real tomogram's own segmentation — per-class intensities measured from the volume itself, tunable object populations, membrane bands, PSF.
• The noise is a parametric spectral model (anisotropy, banding, axial correlation) fitted to background crops of the same stack by a contamination-tolerant estimator.

The claim is deliberately realism, not recreation: the pairs are not reproductions of a specific stack's noise, but stacks with realistic noise and structure — close enough that behavioral conclusions transfer.

Validation, Not Assertion

A three-claim protocol makes realism falsifiable:

Claim	Test	Headline result
C1 — Noise integrity	Purity test separating clean from contaminated noise sources	p < 10⁻⁵
C2 — Structural integrity	Corruption suite of 45 deliberate defects	38/45 harmful corruptions detected
C3 — Behavioral realism	Denoisers must relate to the pair as they relate to the real stack	T1 residual behavior, T2 training transfer (within 2 dB), T3 ranking transfer (88% of pairwise orderings vs 63% under white noise)

Every test ships a negative control that must fail: white noise at the correct amplitude is rejected in 30/30 cells, and a reduced axially-uncorrelated model in 27/30. Acceptance criteria were frozen before the tiers ran, and the scoreboard is reported failure-inclusive. On a seventh, held-out volume, denoiser rankings transfer through the pair (Kendall τ +0.67) while a white-noise benchmark nearly inverts them (τ −0.71).

The benchmark pairs: real slice, clean phantom, and noisy phantom for each volume

What You Get

• Seven published pairs (six development + one held-out) with clean/noisy stacks, labels, frozen configs, background crops, fitted noise models, and evaluation windows — on Zenodo, DOI 10.5281/zenodo.22084921 (CC-BY-4.0)
• The renderer: an interactive tuning workbench plus headless tools to build a matched phantom from your own tomogram + segmentation + background crop
• The validation scripts that scored the published pairs — rerunnable on yours

Documentation

Getting Started

Install the renderer and open your first volume

Tuner Workbench

The interactive phantom-building UI

Tutorial: Your Own Volume

From tomogram + labels + crop to a matched pair

The Benchmark Data

The seven pairs, sources, and licences

Concepts

Concept	Description
Realism, Not Recreation	What PhantEM claims, and what it deliberately does not
The Noise Model	Spectrum, banding, and axial correlation, fitted from background crops
The Validation Protocol	C1/C2/C3, frozen criteria, negative controls

Availability

• Code: github.com/lucasfortune/phantem (BSD-3-Clause, results at tag v1.1)
• Data: Zenodo 10.5281/zenodo.22084921 (CC-BY-4.0)
• Paper: preprint link coming soon
• PhantEM is the benchmark behind the AutoStructN2V evaluation