Segmentation Module

Train U-Net models for biomedical image segmentation or use pretrained models to segment new data.

The Segmentation module showing the step navigation and workflow selection

---

Overview

The Segmentation module provides a complete machine learning pipeline for training and applying U-Net segmentation models. It supports two workflows:

Workflow	Description	Steps
Train from Scratch	Train a new model using your images and annotations	Step 1 → Step 2 → Step 3 → Step 4
Import Pretrained Model	Use an existing model to segment new data	Step 1 → Step 4 (skip Steps 2-3)

For more background, see Segmentation Overview.

---

Quick Start

Training Workflow

1. Launch the Segmentation module from the Hub
2. Expand "Train from Scratch" and select your raw images and annotations
3. Configure training parameters (or use defaults)
4. Start training and monitor progress
5. Upload inference data and run segmentation

Import Model Workflow

1. Launch the Segmentation module from the Hub
2. Expand "Use Pretrained Model" and select your model (.pth) and config (.json) files
3. Skip to Step 4, upload inference data, and run segmentation

---

Step-by-Step Guide

Step 1: Training Data or Model Selection

This step offers two mutually exclusive workflows. Expanding one section will collapse the other. For help choosing, see Workflow Choice.

Choose between training a new model or importing an existing one

Option A: Train from Scratch

Expand the Train from Scratch section to upload your training data:

Input	Format	Description
Raw Images	TIFF stack	Your microscopy images for training. See Raw Images
Annotations	TIFF stack	Corresponding segmentation masks (same dimensions). See Annotations

Both files must have matching dimensions (width, height, number of slices).

To select files:

1. Click the file selector for each input type
2. Choose a file from your workspace or upload a new one
3. Wait for validation to complete (green checkmark indicates success)
4. Click Next: Configuration when both files are validated

Option B: Use Pretrained Model

Expand the Use Pretrained Model section to import an existing model:

Input	Format	Description
Model File	.pth	PyTorch model weights. See Model File
Config File	.json	Model configuration (architecture, classes). See Config File

Note: When using a pretrained model, Steps 2 and 3 are skipped. After validation, click Next to proceed directly to Step 4 (Inference).

---

Step 2: Training Configuration

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

Configure how your model will be trained. Default values work well for most cases.

Training configuration with dataset, architecture, and training parameters

Dataset Configuration

Parameter	Default	Options	Help Article
Patch Size	64	32, 48, 64, 96, 128, 256	Patch Size
Patches per Image	50	1–100	Patches per Image
Batch Size	8	1, 2, 4, 8, 16, 32	Batch Size
Data Augmentation	Enabled	On/Off	Augmentation

Model Architecture

Parameter	Default	Options	Help Article
Number of Features	64	32, 48, 64, 96, 128	Number of Features
Number of Layers	4	2–6	Number of Layers

Training Parameters

Parameter	Default	Options	Help Article
Learning Rate	1e-3	1e-5 to 2e-3	Learning Rate
Number of Epochs	100	10–500	Number of Epochs

Click Next: Training when you're satisfied with your configuration.

---

Step 3: Training Progress

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

Monitor your model as it learns to segment your images. For details on interpreting training metrics, see Training Progress.

Training progress with real-time metrics and charts

Starting Training

1. Review your configuration summary
2. Click Start Training to begin
3. The progress display will show:
   • Current epoch / total epochs
   • Progress bar
   • Real-time metrics (Training Loss, Validation Loss, Training Dice, Validation Dice)

Training Charts

Chart	Description	Help Article
Loss Curves	Shows training and validation loss over epochs	Loss Curves
Dice Score	Shows segmentation accuracy over epochs	Dice Score

Good training shows both curves converging, with validation metrics close to training metrics.

During Training

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 training session can run at a time.

• Cancel Training: Click Cancel Training if you need to stop early
• Page Refresh: If you accidentally refresh the page, the module will detect an active training session and offer to resume monitoring

Training Completion

When training finishes:

• The Next: Inference button becomes enabled
• Your trained model is automatically saved
• Metrics display shows final values

Click Next: Inference to proceed.

---

Step 4: Run Inference

Apply your trained (or imported) model to segment new data. See Inference Overview.

Upload data and run segmentation with your model

Select Inference Data

1. Click the file selector to choose your inference data
2. Select a TIFF stack from your workspace or upload a new one
3. The data should be similar to your training images (same modality, resolution)

For data requirements, see Inference Data.

Run Segmentation

1. Click Run Segmentation after selecting valid data
2. Monitor the progress bar showing slice-by-slice processing
3. Wait for the completion message

After Completion

When segmentation finishes, you'll see two options:

Action	Description
Open in Image Viewer	View your segmentation results immediately. See the Image Viewer Guide for navigation controls
Start New Analysis	Reset the module to begin a new training or inference workflow

---

Output Files

The Segmentation module saves files to your workspace:

Workflow	Output Location	Files
Training	models/<session>/<training-id>/	best_model.pth, config.json, results.json
Inference	results/<inference-id>/	Segmented TIFF stack, metadata JSON

These files appear in your workspace file browser after processing completes.

---

Troubleshooting

Issue	Possible Cause	Solution
Validation fails for training data	Dimension mismatch between images and annotations	Ensure both TIFF stacks have identical dimensions
Training loss not decreasing	Learning rate too high or data issue	Try reducing learning rate; verify annotations are correct
Poor segmentation quality	Insufficient training data or epochs	Add more training data or increase epoch count
"Model file invalid" error	Incompatible model architecture	Ensure model and config files are from the same training session
Inference very slow	Large image stack	This is normal; processing is slice-by-slice

---

Related Help Articles

Module Overview

• Segmentation Overview

Step 1: Data Selection

• Data Upload
• Workflow Choice
• Raw Images
• Annotations
• Model File
• Config File

Step 2: Configuration

• Patch Size
• Patches per Image
• Batch Size
• Augmentation
• Number of Features
• Number of Layers
• Learning Rate
• Number of Epochs

Step 3: Training

• Training Progress
• Loss Curves
• Dice Score

Step 4: Inference

• Run Inference
• Inference Data