AutoStructN2V

Automated structural noise discovery and routed self-supervised denoising for volumetric electron microscopy.

The Problem

Self-supervised denoising methods like Noise2Void work beautifully for random, pixel-independent noise. But many real microscopy volumes contain structured noise — spatially correlated patterns such as:

• Scan lines and streaks from electron microscopy
• Periodic stripes from camera sensors and detector readout
• Directional artifacts from tomographic reconstruction

For structured noise, the core assumption of N2V breaks down: neighboring pixels share correlated noise, and the network can "cheat" by copying it from context instead of removing it. Structured Noise2Void (StructN2V) fixes this with a structural mask — but that mask traditionally has to be specified by hand, from expert reading of autocorrelation plots.

The Solution

ASN2V replaces the expert with a measurement:

1. Measure — the noise autocorrelation (ACF) is measured directly on the raw volume through automatic background selection. Takes seconds, no training.
2. Route — a calibrated routing decision, taken before any training, determines whether structured masking is warranted at all (directionality statistic Dmax against a validated threshold).
3. Extract — when it is, the mask is discovered under explicit design rules: sign-agnostic one-pixel-wide line summaries of the significant ACF features, with an effect-size floor.
4. Train once — exactly one model is trained: StructN2V with the discovered mask, or plain N2V when the noise carries no usable directional structure (the same mechanism with a single-pixel mask).

from autoStructN2V.pipeline import run_pipeline

summary = run_pipeline({
    "input_data": "/path/to/noisy_stack.tif",
    "output_dir": "./results",
    "mask": {"source": "extractor",
             "extractor": {"bg_side": "light"}},
})
# summary["branch"], summary["route_reason"], summary["denoised_stack"]

Key Features

• Self-supervised — no clean reference images required
• Measurement before training — the mask and routing decision are available in seconds, on CPU, before any GPU time is spent
• Safe fallback — noise without usable directional structure routes to plain N2V automatically; abstaining is an informative outcome, not an error
• One training run — no multi-stage orchestration; cost comparable to plain N2V
• Validated — on the PhantEM benchmark, the discovered masks match an oracle mask built from the true noise under identical design rules

Documentation

Getting Started

Installation and the two quickstarts

Basic Tutorial

Complete walkthrough on benchmark data

Pipeline Guide

Configure and run the routed workflow

API Reference

Complete module documentation

Examples

Real-world application examples

Concepts

Understanding the key ideas behind autoStructN2V:

Concept	Description
The Routed Pipeline	Measure → route → train exactly one model
Noise Measurement & Routing	Automatic background selection, the ACF, and the Dmax gate
Spine Mask Extraction	Sign-agnostic line summaries with an effect-size floor
Architecture & Training Recipe	The N2V2-style U-Net and the frozen publication recipe

Package Organization

Component	Purpose
masking/autoextract/	The method proper: measurement, router, spine mask extraction
pipeline/	Routed runner, configuration, data splitting
models/	Network architectures (FlexibleUNet, N2V2 options)
datasets/	Patch sampling with blind-spot masking
trainers/	Training loop with early stopping and logging
inference/	Patch-based prediction for full stacks

Availability

• Code: github.com/lucasfortune/asn2v (BSD-3-Clause, results tagged v1.0)
• No installation needed: ASN2V also runs in the browser as part of the BioMed Workspace — upload a stack and get the full routed run through the same vendored v1.0 code path
• Benchmark data: PhantEM on Zenodo (CC-BY-4.0)
• Paper: preprint link coming soon