AutoStructN2V

A two-stage self-supervised deep learning framework for microscopy image denoising that handles both random and structured noise patterns.

The Problem

Traditional self-supervised denoising methods like Noise2Void work beautifully for random, pixel-independent noise. But many real microscopy images contain structured noise - patterns that repeat spatially across the image:

• Scan lines from electron microscopy
• Periodic stripes from camera sensors
• Wave-like artifacts from scanning mechanisms
• Fixed-pattern noise from detector imperfections

For structured noise, the core assumption of N2V breaks down. The network can "cheat" by learning the noise pattern from context rather than the true signal.

The Solution

AutoStructN2V solves this by treating random and structured noise as separate challenges:

Stage 1: Standard Noise2Void

• Removes random, pixel-independent noise (shot noise, read noise, thermal noise)
• Uses classic blind-spot masking with random pixel selection

Stage 2: Structured Noise2Void

• Removes correlated, pattern-based noise (scan lines, periodic artifacts)
• Uses automatically-discovered spatial masks that target the noise pattern

from autoStructN2V.pipeline import run_pipeline

results = run_pipeline({
    'input_dir': './noisy_images/',
    'output_dir': './results/',
    'run_stage1': True,  # Random noise
    'run_stage2': True   # Structured noise
})

Key Features

• Self-supervised - No clean reference images required for training
• Two-stage pipeline - Comprehensive handling of random + structured noise
• Automatic pattern discovery - Autocorrelation analysis finds noise patterns
• Flexible U-Net - Resize convolution prevents checkerboard artifacts
• ROI selection - Intelligent patch sampling improves training efficiency
• 2.5D mode - Volumetric data support using consecutive slice triplets

Documentation

Getting Started

Installation and quick start guide

Basic Tutorial

Complete beginner walkthrough

Pipeline Guide

Configure and run the full workflow

API Reference

Complete module documentation

Examples

Real-world application examples

Concepts

Understanding the key ideas behind autoStructN2V:

Concept	Description
Two-Stage Approach	How Stage 1 and Stage 2 work together
Structural Mask Extraction	Automatic pattern discovery via autocorrelation
Architecture	Flexible U-Net with resize convolution
ROI Selection	Intelligent patch sampling strategy

Module Organization

The autoStructN2V package is organized into six core components:

Component	Purpose
models/	Neural network architectures (FlexibleUNet)
datasets/	Data loading with masking, augmentation, and ROI selection
masking/	Structural noise extraction and mask generation
trainers/	Training loops with early stopping and logging
inference/	Patch-based prediction for full images
pipeline/	Orchestration of the complete two-stage workflow