Cell Segmentation Guide¶

This guide helps you choose the optimal segmentation method for your bacterial images and achieve the best possible cell detection results.

🎯 Overview¶

mAIcrobe offers four main segmentation approaches:

Although StarDist and U-Net models require custom training, mAIcrobe includes a pre-trained models for several bacterial species and modalities. The models are downloaded from our GitHub repository on first use and stored inside your user folder under .maicrobecache.

Tip: You can always refine the segmentation manually using napari’s built-in label editing tools. Check the official napari Labels layer documentation for more details on using Labels layers.

🌟 StarDist Models¶

Best for: Star-convex shaped cells (most bacteria)

Key Features¶

• 🎯 Purpose: Deep learning-based segmentation for star-convex shapes
• 🔧 Requirement: Custom trained model needed.

Getting Started¶

1. Learn more: Check the StarDist paper and repository
2. Training: Use our example notebook at notebooks/StarDistSegmentationTraining.ipynb Notebook link: https://github.com/HenriquesLab/mAIcrobe/blob/main/notebooks/StarDistSegmentationTraining.ipynb
3. Examples: See StarDist training examples

Note: mAIcrobe provides a pre-trained StarDist model for S. aureus SIM images stained with NileRed membrane dye. Select "StarDist S.aureus SIM" in the segmentation widget.

🔬 Cellpose Models¶

Best for: General cell segmentation across diverse cell types

Key Features¶

• 🎯 Purpose: Universal deep learning segmentation model
• 🚀 Ready to use: Pre-trained cyto3 model included
• 🌐 Versatile: Trained on diverse cell types and imaging modalities

Getting Started¶

1. Learn more: Check the Cellpose paper and repository
2. First run: Model weights download automatically on first use (can take several minutes depending on your internet connection)
3. Usage: Select "CellPose cyto3" in the segmentation widget

Tip: Cellpose is great for getting started quickly without training custom models.

🧠 U-Net Models¶

Best for: Custom applications with specific imaging conditions

Key Features¶

• 🎯 Purpose: Convolutional neural network
• 🔧 Format: Requires Keras model files (.keras)
• 🎨 Flexible: Can be trained for specific cell types and conditions
• ⚙️ ZeroCostDL4Mic: Recommended tool for training U-Net models that integrate with mAIcrobe

Model Requirements¶

Your U-Net model should output:

• 0: Background
• 1: Cell boundary
• 2: Cell interior

mAIcrobe converts this to individual cell labels using scikit-image watershed segmentation.

Note: mAIcrobe provides several pre-trained U-Net models for different bacterial species and imaging modalities. In the compute_label widget, select from the following options:

• "Ph.C. S. pneumo" : Phase contrast S. pneumoniae
• "WF FtsZ B. subtilis": Widefield fluorescence B. subtilis expressing FtsZ-GFP
• "Unet S. aureus": Membrane labeled SIM S. aureus

Getting Started¶

1. Learn more: Read the U-Net paper
2. Training: Use ZeroCostDL4Mic
3. Technical details: See watershed documentation

⚡ Thresholding-Based Methods¶

Best for: Quick analysis without training requirements

Key Features¶

• 🚀 Speed: Fastest segmentation method
• 🔧 No training: Classical image processing
• ⚖️ Trade-off: Lower accuracy for complex images

Available Methods¶

📊 Isodata Thresholding¶

• Type: Global automatic threshold
• How it works: Analyzes image histogram to find optimal threshold
• Best for: Images with clear intensity separation. Phase contrast is a typical example.
• Reference: scikit-image documentation

🎯 Local Average Thresholding¶

• Type: Adaptive local threshold
• How it works: Computes threshold based on local neighborhood
• Best for: Images with uneven illumination
• Reference: scikit-image documentation

Processing Pipeline¶

1. Threshold → Binary image
2. Distance transform → Separate touching cells. See scipy distance transform docs
3. Watershed → Individual cell labels. See scikit-image watershed docs.

📏 Validation and Quality Control¶

✅ Manual Validation Checklist¶

Always validate segmentation results:

• Sample size: Check 50-100 cells randomly
• Visual inspection: Look for common segmentation errors:
  • Under-segmentation (multiple cells as one)
  • Over-segmentation (one cell split into multiple)
  • Boundary accuracy
  • Missing cells

📊 Automated Quality Metrics¶

Key indicators to monitor:

• Cell count consistency across similar images
• Size distribution of segmented cells - look for outliers
• Circularity of segmented cells - should be species-appropriate |

📚 Further Reading¶

• Cell Analysis Guide - What to do after segmentation
• Cell Classification Guide - Cell classification workflows
• API Reference - Programmatic control

📖 Scientific References¶

• StarDist: Schmidt et al., MICCAI 2018
• Cellpose: Stringer et al., Nature Methods 2021
• U-Net: Ronneberger et al., MICCAI 2015

🛠️ Technical Documentation¶

• Watershed segmentation: scikit-image docs
• Image filters: scikit-image docs

Next: Learn how to analyze your segmented cells in the Cell Analysis Guide 🔬