Deep Learning Denoising Module

Remove noise from microscopy images using self-supervised deep learning. No clean reference images required.

The DL Denoising module showing method selection, mode toggle, and workflow options

---

Overview

The Deep Learning Denoising module provides two self-supervised denoising methods:

Method	Training	Best For	Complexity
N2V (Noise2Void)	Single stage	Random, uncorrelated noise (Gaussian, Poisson)	Simple
autoStructN2V	Two stages + mask extraction	Structured noise (scan lines, periodic artifacts, camera patterns)	Advanced

Both methods learn to denoise from noisy images alone—no clean ground truth required.

For method selection guidance, see Denoising Method.

When to Use Each Method

Your Images Have...	Recommended Method
Random speckle noise	N2V
Poisson (shot) noise	N2V
Gaussian noise	N2V
Horizontal/vertical scan lines	autoStructN2V
Periodic stripe artifacts	autoStructN2V
Camera-specific fixed patterns	autoStructN2V
Tomography reconstruction artifacts	autoStructN2V

---

Quick Start

Decision Flowchart

Start
  │
  ├─ Do you have a pretrained model?
  │     │
  │     ├─ YES → Import Model workflow (skip training)
  │     │
  │     └─ NO → Continue
  │
  ├─ Does your noise have visible patterns or structure?
  │     │
  │     ├─ YES → Use autoStructN2V
  │     │
  │     └─ NO (random noise) → Use N2V
  │
  ├─ Is noise correlated between slices? (volumetric data)
  │     │
  │     ├─ YES → Use 2.5D mode (requires 20+ slices)
  │     │
  │     └─ NO → Use 2D mode
  │
  └─ Configure → Train → Done

N2V Quick Path

1. Launch module → Select N2V → Keep 2D mode
2. Choose Train from Scratch → Upload your noisy TIFF stack
3. Use Balanced preset (or customize)
4. Click Start Denoising → Wait for completion
5. View results in Image Viewer

autoStructN2V Quick Path

1. Launch module → Select autoStructN2V → Choose 2D or 2.5D mode
2. Choose Train from Scratch → Upload your noisy TIFF stack
3. Use Balanced preset (or customize both stages)
4. Click Start Denoising
5. After Stage 1: Review the extracted mask → Approve & Continue
6. Wait for Stage 2 → View results

---

Step-by-Step Guide

Step 1: Method & Data Selection

Step 1 guides you through three decisions: method, mode, and workflow.

Select your denoising method, toggle 2D/2.5D mode, then choose a workflow

1. Select Denoising Method

Choose between the two methods at the top of the step:

Option	Description
Noise2Void (N2V)	Fast single-stage training. Best for random noise.
autoStructN2V	Two-stage training with automatic structured noise analysis. Best for patterned artifacts.

See Denoising Method for detailed guidance.

2. Select Processing Mode (2D / 2.5D)

Use the toggle switch in the top-right corner of the method selection card:

Mode	Description	Requirements
2D (default)	Process each slice independently	Any stack size
2.5D	Process triplets (Z-1, Z, Z+1) using inter-slice context	Minimum 20 slices

When to use 2.5D mode:

• Volumetric data with correlated structures across slices
• EM tomography and confocal microscopy stacks
• Noise patterns that span multiple slices
• When preserving 3D structural continuity is important

2.5D requirements:

• Minimum 20 slices in the stack
• Slices should be spatially contiguous (not random samples)
• More GPU memory required (3x input channels)

Important for autoStructN2V: In 2.5D mode, the mask extraction phase shows a 3-slice triplet mask instead of a single 2D mask. See Step 3: Mask Extraction (2.5D) for details.

See 2D vs 2.5D Processing Mode for more details.

3. Choose Workflow

After selecting a method, two workflow options appear:

Option A: Train from Scratch

Expand Train from Scratch to upload your data:

1. Click the file selector
2. Choose a TIFF stack from your workspace or upload a new file
3. Wait for validation (green checkmark)
   • For 2.5D mode: validation checks that the stack has at least 20 slices
4. Click Next: Configure

See Input Data for file requirements.

Option B: Import Previously Trained Model

Expand Import Previously Trained Model to use an existing model:

For N2V:

File	Description
Config (.json)	Training configuration. See Config File
Model (.pth)	Trained model weights. See Model File

For autoStructN2V:

File	Description
Config (.json)	Training configuration
Stage 1 Model (.pth)	Stage 1 trained weights
Stage 2 Model (.pth)	Stage 2 trained weights

Note: When importing a model, Steps 2 and 3 are skipped. Proceed directly to Step 4 (Inference).

---

Step 2: Configure Training

This step is only shown for the Train from Scratch workflow.

Configure your training parameters. The interface differs based on your selected method.

Presets

Start with a preset to set all parameters at once:

Preset	Description
Fast	Quick training, lower quality. Good for testing.
Balanced	Recommended for most cases. Good quality with reasonable training time.
High Quality	Best results, longer training time. Use when quality is critical.

N2V Configuration (Single Column)

For N2V, you'll see a single configuration form:

N2V single-stage configuration

Parameter Summary

Category	Parameter	Default	Help Article
Dataset	Patch Size	64	Patch Size
	Patches per Image	100	Patches per Image
	Batch Size	8	Batch Size
	Mask Percentage	15%	Mask Percentage
	Data Augmentation	On	Augmentation
Architecture	Number of Features	64	Features
	Number of Layers	4	Number of Layers
Training	Learning Rate	1e-4	Learning Rate
	Epochs	100	Epochs
	Early Stopping	On	Early Stopping

Advanced Options (collapsed by default):

• Resize Convolution: Resize Conv
• Upsampling Mode: Upsampling Mode
• Masking Strategy: Masking Strategy

autoStructN2V Configuration (Dual Column)

For autoStructN2V, you'll see a side-by-side layout for both training stages:

autoStructN2V two-stage configuration with Stage 1 (left) and Stage 2 (right)

Stage 1 vs Stage 2 Defaults

Parameter	Stage 1	Stage 2	Notes
Patch Size	32	64	Stage 2 uses larger patches
Patches per Image	100	200	More patches in Stage 2
Batch Size	4	2	Smaller batches in Stage 2
Mask Percentage	15%	10%	Lower masking in Stage 2
Learning Rate	1e-4	1e-5	Lower learning rate in Stage 2
Data Augmentation	Off	On	Augmentation disabled in Stage 1

Important: Stage 1 augmentation is automatically disabled for autoStructN2V. This preserves structural noise patterns needed for mask extraction.

Mask Extractor Configuration

Below the stage columns, expand Mask Extractor Configuration to adjust how structural noise patterns are detected:

Parameter	Default	Description	Help Article
Adaptive Thresholding	On	Automatically adjust threshold per slice	Adaptive Thresholding
Base Percentile	50%	Starting threshold for noise detection	Base Percentile
Percentile Decay	1.15	Rate at which threshold adapts	Percentile Decay
Max Masked Pixels	25	Maximum active pixels in the mask	Max Masked Pixels

Click Next: Training when configuration is complete.

---

Step 3: Training

This step is only shown for the Train from Scratch workflow.

The training interface differs significantly between N2V and autoStructN2V.

N2V Training (Single Stage)

For N2V, click Start Denoising to begin training:

N2V single-stage training progress

During Training:

• Progress bar shows current epoch
• Loss chart displays training and validation loss
• Metrics cards show current and best validation loss

Training Completion:

• Status changes to "Complete"
• Open in Viewer and Start new Analysis buttons appear
• You can proceed to Step 4 to denoise additional images

Tip: You don't need to watch the entire training process. Feel free to explore other modules while training continues in the background. Note that only one denoising training can run at a time.

autoStructN2V Training (Multi-Stage)

autoStructN2V training consists of three phases displayed in collapsible sections:

autoStructN2V three-phase training interface

Phase 1: Stage 1 N2V Training

Click Start Denoising to begin. Stage 1 trains a standard N2V model:

• Progress bar shows epoch progress
• Loss chart tracks training/validation loss
• Status updates: "Pending" → "Training" → "Complete"

When Stage 1 completes, the mask extraction phase begins automatically.

Phase 2: Mask Extraction

After Stage 1, the module analyzes the difference between noisy input and Stage 1 output to extract a structural noise mask.

Mask Extraction (2D Mode)

In 2D mode, you'll see a single mask grid:

2D mask extraction showing a single detected noise pattern

Mask Display:

• Single grid showing the detected structural noise pattern
• Active pixels (purple) represent positions where structured noise was detected
• Center pixel (highlighted) is the reference position
• Statistics: kernel size, active pixel count, coverage percentage

Mask Extraction (2.5D Mode)

In 2.5D mode, the mask becomes a triplet showing patterns for three Z-positions:

2.5D triplet mask extraction with three slice tabs

Triplet Mask Interface:

• Three tabs at the top: "Z-1 (above)", "Z (center)", "Z+1 (below)"
• Each tab shows the mask for that slice position in the triplet
• Click tabs to switch between slice views
• The 3D mask captures noise correlations across the Z-axis

Interpreting the Triplet Mask:

Tab	Shows	Meaning
Z-1 (above)	Mask for slice above center	Noise correlations from previous slice
Z (center)	Mask for center slice	Primary structural noise pattern
Z+1 (below)	Mask for slice below center	Noise correlations from next slice

Different slices may show different patterns if the structured noise varies across the Z-axis.

Mask Actions

Button	Description
Approve & Continue to Stage 2	Accept the mask and proceed to Stage 2 training
Skip Stage 2	Use Stage 1 results only (if no structure detected)

Auto-approve Toggle: Enable Auto-approve (toggle in header) to automatically approve the mask and continue to Stage 2 without manual intervention. Useful for batch processing.

Adjusting the Mask: If the detected pattern doesn't look right, use the parameter panel to adjust:

1. Modify Base Percentile, Percentile Decay, or Max Masked Pixels
2. Click Regenerate Mask
3. Review the new mask (check all three tabs in 2.5D mode)
4. Approve when satisfied

Phase 3: Stage 2 Struct-N2V Training

After mask approval, Stage 2 trains using the extracted mask:

• Progress and metrics display similar to Stage 1
• Stage 2 specifically targets the structured noise patterns identified in the mask
• In 2.5D mode, the 3D mask guides the network to ignore correlated noise across slices
• Status: "Pending" → "Training" → "Complete"

When Stage 2 Completes:

• Open in Viewer button appears
• Final denoised result is the Stage 2 output
• You can compare Stage 1 vs Stage 2 results in the Image Viewer

---

Step 4: Process Additional Data (Optional)

Apply your trained model to denoise additional images without retraining.

Apply the trained model to new data

Model Information

A summary card shows your model details:

• Method used (N2V or autoStructN2V)
• Training mode (2D or 2.5D)
• When training completed

Process New Images

1. Click the file selector to choose additional data
2. Select a TIFF stack from your workspace
   • For 2.5D models: new data should also have sufficient slices
3. Click Process Data
4. Monitor the progress bar
5. When complete, click Open in Image Viewer or Process More

See Inference Data and Inference Overview.

---

Output Files

The module saves files to your workspace:

Output	Location	Description
Denoised images	results/<training-id>/	Denoised TIFF stack
Trained model	models/<session>/<training-id>/	Model weights (.pth)
Configuration	models/<session>/<training-id>/	Training config (.json)
Mask (autoStructN2V)	models/<session>/<training-id>/	Extracted noise mask

For autoStructN2V, both Stage 1 and Stage 2 outputs are saved.

---

Troubleshooting

Issue	Possible Cause	Solution
"Insufficient Stack Depth for 2.5D"	Stack has fewer than 20 slices	Switch to 2D mode or use a larger stack
Training very slow	Large images or many epochs	Reduce patch size, use Fast preset, or wait
Poor denoising quality	Wrong method for noise type	Try the other method (N2V ↔ autoStructN2V)
Artifacts remain after N2V	Structured noise present	Switch to autoStructN2V
Empty mask extracted	No detectable structure	Use Skip Stage 2 or try N2V instead
Mask shows random pattern	Threshold too sensitive	Increase Base Percentile, reduce Max Masked Pixels
2.5D triplet mask looks inconsistent	Noise varies across slices	This may be correct; check each tab individually
Stage 2 worse than Stage 1	Mask captured wrong pattern	Adjust mask parameters and regenerate
GPU not detected	CUDA not available	Training will use CPU (slower)
Out of memory	Batch size too large	Reduce batch size (especially in 2.5D mode)

---

Method Details: N2V

Noise2Void (N2V) is a self-supervised denoising method that works by:

1. Blind-spot training: During training, the network learns to predict each pixel's value from its surrounding context, without seeing the pixel itself
2. Self-supervision: The noisy image serves as both input and target
3. Assumption: Noise is pixel-independent (random, uncorrelated)

In 2.5D mode: The network takes 3 consecutive slices as input and predicts only the center slice, leveraging inter-slice context for better denoising.

Strengths:

• Fast single-stage training
• Works well for most microscopy noise
• Simple configuration

Limitations:

• Cannot handle structured or correlated noise
• May leave pattern artifacts from camera or reconstruction

For more details, see N2V Training.

---

Method Details: autoStructN2V

autoStructN2V extends N2V to handle structured noise through a two-stage approach:

Stage 1: Initial N2V Training

• Trains a standard N2V model
• Output: Partially denoised image (random noise removed, structured noise remains)

Mask Extraction:

• Compares noisy input with Stage 1 output
• Identifies pixels with consistent residual patterns
• Generates a "structN2V mask" showing which pixel relationships to exclude
• In 2.5D mode: Generates a 3D triplet mask capturing patterns across Z-1, Z, and Z+1 positions

Stage 2: Struct-N2V Training

• Trains using the extracted mask
• The mask tells the network which neighboring pixels might share correlated noise
• In 2.5D mode: The 3D mask guides cross-slice noise handling
• Output: Fully denoised image (both random and structured noise removed)

Strengths:

• Automatically detects structured noise patterns
• No manual mask creation required
• Effective for scan lines, periodic artifacts, camera patterns
• 2.5D mode captures 3D noise correlations

Limitations:

• Longer training time (two stages)
• Requires sufficient structural noise for mask extraction
• More parameters to configure

For more details, see autoStructN2V Training and autoStructN2V Details.

---

Related Help Articles

Module Overview

• DL Denoising Overview

Step 1: Method & Data Selection

• Denoising Method
• 2D vs 2.5D Processing Mode
• Workflow Selection
• Input Data
• Import Config File
• Import Model File

Step 2: Configuration

• Configuration Overview
• Patch Size
• Patches per Image
• Batch Size
• Mask Percentage
• Augmentation
• Features
• Number of Layers
• Learning Rate
• Epochs
• Early Stopping

Step 2: Mask Extractor (autoStructN2V)

• Adaptive Thresholding
• Base Percentile
• Percentile Decay
• Max Masked Pixels

Step 3: Training

• Training Overview
• N2V Training
• autoStructN2V Training
• Loss Metrics
• Best Validation Loss
• autoStructN2V Details

Step 4: Inference

• Inference Overview
• Inference Data