Skip to content

usage-guide.md

Latest commit

 

History

History
55 lines (41 loc) · 1.9 KB

usage-guide.md

File metadata and controls

55 lines (41 loc) · 1.9 KB

Cell Segmentation - Usage Guide

Overview

Cell segmentation identifies individual cells in multiplexed images, providing cell masks essential for downstream single-cell analysis.

Prerequisites

pip install cellpose deepcell scikit-image napari
# For GPU acceleration (optional)
pip install cellpose[gpu]

Quick Start

Tell your AI agent what you want to do:

  • "Segment cells in my IMC images using Cellpose"
  • "Run Mesmer segmentation on my tissue images"
  • "Create cell masks from nuclear and membrane channels"

Example Prompts

Cellpose Segmentation

"Segment cells using Cellpose with my DNA channel as nuclear marker"

"Run Cellpose cyto2 model on my IMC image with diameter 30 pixels"

Mesmer/DeepCell Segmentation

"Use Mesmer to segment my tissue image with nuclear and membrane channels"

"Run DeepCell whole-cell segmentation on my multiplexed image"

Parameter Tuning

"Adjust Cellpose flow threshold to reduce oversegmentation"

"Optimize cell diameter parameter for my tissue type"

Quality Control

"Overlay my segmentation mask on the original image for QC"

"Check my segmentation for oversegmentation and undersegmentation"

What the Agent Will Do

  1. Load preprocessed multichannel image
  2. Select appropriate nuclear channel (DNA/Ir-191/193 or histone markers)
  3. Optionally select membrane channel (CD45, Na/K-ATPase)
  4. Run deep learning model (Cellpose or Mesmer)
  5. Post-process masks (remove small objects, fill holes)
  6. Save segmentation mask as labeled image

Tips

  • Measure average cell diameter from representative cells for Cellpose
  • Cellpose models: nuclei (nuclear only), cyto/cyto2 (whole cell)
  • Mesmer is trained specifically on tissue images with better tissue context
  • flow_threshold (0.4 default) controls cell separation
  • cellprob_threshold (0.0 default) controls detection sensitivity
  • Always visually QC segmentation overlaid on original image