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¶

napari-mAIcrobe offers four main segmentation approaches:

Method	Type	Training Required	Speed	Accuracy
⭐ StarDist	Deep Learning	✅ Custom model	Medium	High
🔬 Cellpose	Deep Learning	❌ Pre-trained	Medium	High
🧠 U-Net	Deep Learning	✅ Custom model	Medium	High
⚡ Thresholding	Classical	❌ None	Fast	Medium

⭐ StarDist Models¶

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

Key Features - Purpose: Deep learning-based segmentation for star-convex shapes - Performance: High accuracy for bacterial cells - Requirement: Custom trained model needed

Getting Started 1. Train with our example notebook: mAIcrobe/notebooks/StarDistSegmentationTraining.ipynb 2. See StarDist training examples in the official repo

Note: mAIcrobe doesn't include pre-trained StarDist models—you must provide your own.

🔬 Cellpose Models¶

Best for: General cell segmentation across diverse cell types

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

Usage: Select "CellPose cyto3" in the segmentation widget.

🧠 U‑Net Models¶

Best for: Custom applications with specific imaging conditions

Format: Keras model files (.keras)
Model output: 0 background, 1 boundary, 2 interior
mAIcrobe converts to individual cell labels via watershed

Training resources: ZeroCostDL4Mic; Watershed via scikit-image.

⚡ Thresholding-Based Methods¶

Best for: Quick analysis without training requirements

Available Methods - Isodata Thresholding (global) - Local Average Thresholding (adaptive)

Processing Pipeline 1. Threshold → Binary image 2. Distance transform → Separate touching cells 3. Watershed → Individual cell labels

📏 Validation and Quality Control¶

Manual Validation Checklist - Sample size: Check 50–100 cells randomly - Visual inspection: under/over-segmentation, boundary accuracy, missing cells

Automated Quality Indicators - Cell count consistency across similar images (±10%) - Reasonable size distribution, few outliers

🧭 Choosing the Right Method¶

Your Situation	Recommended Method
Quick start, no training	Cellpose cyto3
Very fast, simple images	Isodata thresholding
Best accuracy, have training data	StarDist (custom)
Specific imaging conditions	U-Net (custom)
Uneven illumination	Local average thresholding

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