Cell Classification Guide¶

This guide provides comprehensive information about mAIcrobe's cell classification system, including model selection, usage, and custom model integration.

🧠 Overview¶

mAIcrobe uses deep learning models to automatically classify cells based on their morphological and fluorescence features. The plugin includes:

• 6 pre-trained models optimized for the cell cycle stage detection of Staphylococcus aureus under various imaging conditions
• 1 pre-trained model for E. coli antibiotic phenotyping
• Support for user-trained custom models

🔬 How It Works¶

The classification model is a CNN implemented in TensorFlow/Keras. The architecture was previously described in the eHooke publication. Classification is performed at the single-cell level using cropped images extracted from segmented cells.

For each cell, the following preprocessing steps are applied before feeding it to the model:

• Crops the cell with a small margin (user defined and model dependent but defaults to 5 px).

• Masks the crop by the cell shape defined by the label and rescales intensities to [0, 1].

• Pads the crop with zero to a square shape (final shape model dependent).

• Resizes images to 100×100. If two channels, concatenates the channels side‑by‑side into a 100×200 image.

🔬 Pre-trained Models¶

mAIcrobe includes 7 specialized models optimized for different imaging conditions and channel availability.

🧬🔴 DNA + Membrane Models¶

S.aureus DNA+Membrane Epi¶

• Imaging: Epifluorescence microscopy
• Channels: DNA stain (e.g., Hoechst) + Membrane stain (e.g., NileRed)
• Use case: Cell cycle phase detection in S. aureus in standard fluorescence imaging with both channels

S.aureus DNA+Membrane SIM¶

• Imaging: Structured illumination microscopy (SIM)
• Channels: DNA stain + Membrane stain
• Use case: Cell cycle phase detection in super-resolution imaging

E.coli DNA+Membrane AB phenotyping¶

• Imaging: Epifluorescence microscopy
• Channels: DNA stain + Membrane stain
• Use case: Phenotyping antibiotic-treated E.coli cells. Distinguishes between control cells and those treated with Mecillinam or Nalidixate.

🧬 DNA Only Models¶

S.aureus DNA Epi¶

• Imaging: Epifluorescence microscopy
• Channels: DNA stain only
• Use case: Cell cycle detection in S. aureus when membrane staining is not available/desired
• Accuracy: Lower than dual-channel models or membrane-only models

S.aureus DNA SIM¶

• Imaging: Structured illumination microscopy
• Channels: DNA stain only
• Use case: Cell cycle detection in S. aureus when membrane staining is not available/desired
• Accuracy: Lower than dual-channel models or membrane-only models

🔴 Membrane Only Models¶

S.aureus Membrane Epi¶

• Imaging: Epifluorescence microscopy
• Channels: Membrane stain only
• Use case: Cell cycle detection in S. aureus when DNA staining is not available/desired
• Accuracy: Comparable to dual-channel models, often better than DNA-only models

S.aureus Membrane SIM¶

• Imaging: Structured illumination microscopy
• Channels: Membrane stain only
• Use case: Cell cycle detection in S. aureus when DNA staining is not available/desired
• Accuracy: Comparable to dual-channel models, often better than DNA-only models

Note: Values represent the percentage of samples classified into each category. Diagonal values indicate correct classifications, while off-diagonal values represent misclassifications.

🎨 Custom Model Integration¶

Load and use your own trained TensorFlow models.

📋 Model Requirements¶

Supported Formats: - Keras models (.keras)

🔬 Model Training Guidelines¶

Training Data Requirements:

• Manually annotated cell images
• Balanced representation of all classes
• Consistent imaging conditions

To train your own models and assure seamless integration with the plugin, refer to the jupyter notebook: Cell Cycle Model Training

🥒 Build Your Own Training Data (Pickles)¶

For more detailed information refer to the tutorial: Generate Training Data for Cell Classification¶

Use the Compute pickles widget to export standardized per-cell crops:

🛠️ Workflow:¶

1. Labels layer: Ensure you have a Labels layer (cells). This is used to detect individual cells.
2. Image layers: Ensure you have one or two Image layers. These are used to extract the training crops from.
3. Points layer: Create a Points layer and name it as a positive integer class id (e.g., "1"); add one point per cell to assign that class. Make sure to repeat for other classes with different integer names.
4. Export: Open Plugins > mAIcrobe > Compute pickles, select the layers and output folder, then click "Save Pickle".

⚠️ Important Notes:¶

• The Points layer name must be a positive integer (class id). This is required.
• Each point in the Points layer assigns the corresponding cell to that class. Make sure to add only one point per cell and one cell per class.

💾 What gets saved:¶

• Class_<id>_source.p: list of masked, padded, and resized crops (100×100; 100×200 if two channels concatenated).
• Class_<id>_target.p: list with the class id repeated for each crop.

These files integrate with the training notebook: Cell Cycle Model Training

� Further Reading¶

• Cell Analysis Guide - Complete analysis workflows
• Getting Started - Basic usage tutorial
• API Reference - Programmatic control

Next: Explore programmatic usage in the API Reference.