TUM.ai Neurogenesis Napari Plugin

A napari plugin for automated nuclear segmentation and neural cell type classification in neurogenesis research. Supports classification of astrocytes, neurons, oligodendrocyte precursor cells (OPCs), and dead cells from multi-channel fluorescence microscopy images.

Key Features

Widget	Function	Input	Output
Normalize + Denoise	Color normalization and denoising	Bright-field image	Processed image
Segment	Nuclear segmentation	DAPI/nuclear stain	Masks, centroids, bounding boxes
Segment + Classify	End-to-end cell analysis	4-channel images	Cell segmentation + classification

Quick Start

Installation

pip install neurogenesis-napari

Or install through napari:

1. Open napari
2. Go to Plugins → Install/Uninstall Plugins
3. Search for "TUM.ai Neurogenesis Toolkit"
4. Click Install

Basic Usage

1. Load your images into napari
2. Select the appropriate widget from the Plugins menu
3. Choose your image layers from the dropdown menus
4. Click the action button to process

Model weights are automatically downloaded on first use.

---

Widget Documentation

Normalize + Denoise

Standardizes color variations and reduces noise in bright-field microscopy images.

Usage

1. Load a bright-field image into napari
2. Open Plugins → Normalize and Denoise
3. Select your bright-field image from the BF dropdown
4. Click "Normalize + Denoise"

What it does

• Color Normalization: Histogram matching against an internal reference to standardize appearance across different acquisitions
• Denoising: Removes noise while preserving cellular structures using Cellpose
• Output: Creates a new layer named {original_name}_denoised

---

Segment

Detects and segments individual cell nuclei from DAPI-stained images using Cellpose.

Usage

1. Load a DAPI/nuclear staining image into napari
2. Open Plugins → Segment
3. Select your DAPI image from the DAPI dropdown
4. Optionally adjust:
   • GPU: Enable for faster processing (if CUDA available)
   • Model: Choose Cellpose model (cyto3 default)
5. Click "Segment"

What it does

• Segmentation: Uses Cellpose to identify individual nuclei
• Creates 3 new layers:
  • {name}_masks: Segmentation masks for each nucleus
  • {name}_centroids: Center points of detected nuclei
  • {name}_bboxes: Bounding boxes (polygons) around each nucleus

---

Segment + Classify

End-to-end pipeline that segments nuclei and classifies neural cell types in multi-channel fluorescence images.

Usage

1. Load a 4-channel image into napari as separate layers:
   • DAPI: Nuclear staining (DAPI/Hoechst)
   • Tuj1: β-III-tubulin (neuronal marker)
   • RFP: Red fluorescent protein marker
   • BF: Bright-field
2. Open Plugins → Segment and Classify
3. Select each channel from the respective dropdowns
4. Choose Reuse cached segmentation:
   • True (default): Reuse previous segmentation if available (faster)
   • False: Perform fresh segmentation
5. Click "Segment + Classify"

How it works

1. Segmentation: Cellpose-based nuclear segmentation on DAPI channel
2. Feature extraction: Variational Autoencoder (VAE) extracts features from 4-channel patches around each nucleus
3. Classification: Nearest-centroid classifier assigns cell types based on learned centroids

Output

Creates colored polygon overlays for each detected cell:

• Astrocytes (magenta)
• Neurons (cyan)
• OPCs - Oligodendrocyte Precursor Cells (lime)
• Dead Cells (gray)

The classification results can be manually corrected through an interactive interface. Select any cell and use keyboard shortcuts to reassign its type: Shift+A (Astrocyte), Shift+N (Neuron), Shift+O (OPC), Shift+D (Dead Cell).

---

---

Technical Details

Supported Image Formats

• .czi (Zeiss microscopy files, via napari-czifile2)
• .tiff, .tif
• .png, .jpg

Cell Classification Model

• Feature extraction: Variational Autoencoder (VAE) with 2304-dimensional latent space
• Classifier: Scikit-learn Nearest Centroid
• Input: 224×224 pixel patches from 4 channels (DAPI, BF, Tuj1, RFP)
• Output: 4 cell type classes

Requirements

• Python ≥ 3.10
• CUDA-capable GPU (optional, for faster processing)
• Model weights are automatically downloaded on first use via Hugging Face Hub

---

Citation

If you use this plugin in your research, please cite:

TUM.ai Neurogenesis Napari Plugin
Technical University of Munich
https://github.com/tum-ai/neurogenesis_napari

---

License

MIT License - see LICENSE file for details.