Best Practices

This guide covers data preparation, training recommendations, and resource management to help you get the best results from the workspace.

---

Data Preparation

File Format Requirements

The workspace accepts TIFF stacks (.tif or .tiff) for all processing modules.

Requirement	Specification
Format	Multi-page TIFF stack
Bit Depth	8-bit or 16-bit grayscale
Structure	3D stack (slices × height × width)
Channels	Single channel (grayscale)

File Size Limits:

File Type	Maximum Size
TIFF uploads	200 MB
Workspace ZIP	5 GB
Model files	2 GB

Image Quality

For best results, ensure your images have:

• Proper focus — Blurry images reduce model accuracy
• Sufficient contrast — Images with uniform intensity are rejected
• Consistent acquisition — Use the same microscope settings across your dataset

Files automatically rejected:

• Empty files (0 bytes)
• All-zero pixel values (no data)
• Uniform intensity (no contrast)

---

Training Data Guidelines

Raw Training Images

High-quality training images lead to better models:

Guideline	Why It Matters
Include diverse examples	Model learns to handle variation
Vary object sizes	Handles small and large structures
Include different backgrounds	Robust to imaging variations
Match inference conditions	Similar settings improve accuracy
More images = better	Larger datasets improve generalization

Tip: Include examples of all the variations you expect to encounter in your actual data.

Annotation Masks (for Segmentation)

Annotations tell the model what to segment. Quality directly affects results.

Pixel Value Format:

Value	Meaning
0	Background (not segmented)
1	Class 1
2	Class 2
...	Additional classes

Maximum 10 classes supported.

Annotation Quality Checklist:

• [ ] Consistent labeling — Same structures labeled the same way across all slices
• [ ] Accurate boundaries — Trace object edges carefully
• [ ] Complete labeling — Label ALL instances (unlabeled objects = background)
• [ ] Diverse examples — Include objects at different sizes and positions

Warning: Poor annotations are the #1 cause of poor segmentation results. Invest time in quality labels.

Dimension Matching

For segmentation training, raw images and annotation masks must have:

Dimension	Requirement
Slices	Identical count
Width	Identical pixels
Height	Identical pixels

The system validates this automatically and rejects mismatched pairs.

---

Deep Learning Denoising Data

Minimum Requirements

Mode	Minimum Slices
2D	10 slices
2.5D	20 slices

Noise Matching

For best denoising results:

Factor	Recommendation
Acquisition session	Same session preferred
Microscope/camera	Same equipment
Settings	Same imaging parameters
Sample preparation	Same protocol

Important: A model trained on one type of noise won't effectively remove a different type. If your new images have different noise characteristics, train a new model.

Method Selection

Noise Type	Recommended Method
Random speckle noise	N2V
Poisson (shot) noise	N2V
Gaussian noise	N2V
Horizontal/vertical scan lines	autoStructN2V
Periodic stripe artifacts	autoStructN2V
Camera fixed patterns	autoStructN2V

---

Training Tips

Parameter Recommendations

Start with these values and adjust based on your results:

Parameter	Recommended	Notes
Batch Size	4-8	Reduce if running out of GPU memory
Patch Size	64-256	Larger captures more context but uses more memory
Epochs	100	Early stopping will end training if converged
Augmentation	Enabled	Disable only if image orientation is critical
Learning Rate	Default	Start with preset values

Monitoring Training Progress

Watch these indicators during training:

Indicator	What to Look For
Training Loss	Should decrease over epochs
Validation Loss	Should decrease and stay close to training loss
Loss Gap	Large gap between training/validation = overfitting
Dice Score (segmentation)	Should increase toward 1.0

Signs of Problems:

Symptom	Likely Cause	Solution
Loss not decreasing	Learning rate too high	Try lower learning rate
Loss stuck at high value	Bad training data	Check annotations, add more data
Validation much higher than training	Overfitting	Add augmentation, more data
Very slow training	Large images/batch	Reduce batch size or patch size

Background Training

You don't need to watch training continuously:

• Training continues in the background
• Explore other modules while waiting
• Only one training session can run at a time
• If you refresh the page, you can resume monitoring

---

Resource Management

Critical: Session-Based Storage

Warning: The workspace uses session-based storage. Your files exist only during your current session.

Files are cleared when you:

• Log out
• Close the browser (session timeout)
• Session expires due to inactivity

Always backup your work!

Backup Strategy

Follow this workflow to protect your data:

1. Download Workspace Regularly

When to Backup	Why
End of each session	Preserve all work before logging out
After training completes	Save trained models
After processing results	Keep generated outputs
Before major changes	Safety checkpoint

To download:

1. Click the download icon next to "Workspace" in the sidebar
2. Confirm the download
3. Wait for ZIP to generate (large workspaces take longer)
4. Save the file locally

2. Organize Your Backups

Use descriptive filenames with dates:

workspace_2024-01-15_segmentation-trained.zip
workspace_2024-01-16_denoising-complete.zip

3. Verify Your Backups

After restoring, check that file counts match your expectations.

Workspace ZIP Contents

Your backup includes:

Included	Not Included
All uploaded files	Cache directories
Trained models	Thumbnail cache
Inference results	Slice cache
Processing outputs	Mesh preview cache
Folder structure

Restoring a Workspace

To restore from a backup:

1. Click Upload in the file browser
2. Select "Restore Workspace (ZIP)" from the category dropdown
3. Choose your backup ZIP file
4. Confirm the restore

Warning: Restoring REPLACES your current workspace. All existing files are deleted before restoration.

File Size Management

Keep your workspace running smoothly:

Action	Benefit
Delete unused files	Reduce workspace size
Monitor sidebar stats	Track total size
Download and clear periodically	Start fresh with backup available
Keep under 5 GB	Optimal performance

---

Troubleshooting Common Issues

Issue	Likely Cause	Solution
Upload rejected	File too large or wrong format	Check size limits (200 MB), use TIFF format
Dimension mismatch	Raw and annotation sizes differ	Ensure identical dimensions
Poor segmentation	Insufficient or poor quality training data	Add more diverse, well-annotated data
Training loss not decreasing	Learning rate too high	Try default parameters or lower learning rate
Out of memory	Batch size too large	Reduce batch size
Session expired	Inactivity timeout	Log in again, restore from backup
Slow processing	Large images	Normal behavior; wait for completion
Denoising not effective	Wrong method for noise type	Try N2V ↔ autoStructN2V

---

Summary Checklist

Before starting a project:

• [ ] Data format — TIFF stacks, 8/16-bit grayscale
• [ ] File sizes — Under 200 MB per file
• [ ] Training data — Diverse, well-annotated, matching dimensions
• [ ] Backup plan — Know where you'll save workspace ZIPs

During work:

• [ ] Monitor training — Check loss curves for convergence
• [ ] Backup regularly — Download workspace after significant progress
• [ ] Match conditions — Use similar images for training and inference