SLAP2-Cellpose

Model training and evaluation can be found at https://github.com/AllenNeuralDynamics/FORCEB-kv-refstacks-Cellpose/tree/main

Cellpose Nuclei model retrained with SLAP2 data

This repository contains a version of the Cellpose nuclei model that has been retrained using SLAP2 data. The Cellpose model is a generalist algorithm for cell segmentation.

1a. Create an conda env

conda create -n cellpose python=3.12

conda activate cellpose

1b. For SLAP2 rig only:

1. Create an environment:

   /path/to/python3 -m venv cellpose
   

2. Activate the environment:

   cd to C:\Users\ScanImage\Documents\GitHub\SLAP2-Cellpose
   cellpose\Scripts\activate
   

[Warning] Run either 1a or 1b based on your setup.

2. Install cellpose

pip install -U --no-cache-dir git+https://www.github.com/mouseland/cellpose.git

3. Running cellpose

python run.py --input <input_file_path> --output <output_directory_path>

Example

python run.py --input /Users/caleb.shibu/Downloads/725018_20240326_163614_DMD1_merged.tif --output /Users/caleb.shibu/Desktop/test-cellpose

The output folder would have 2 files flows.tif and masks_pred.tif.

Model comparison before and after retraining:

Cyto2 model gave the highest AUC value for CellProbabilty of 2 and FlowThreshold of 0.5. We used that to train cyto2 model with Voltage Imaging data and the AUC value improved for CellProbabilty of -1.0 and FlowThreshold of 0.5.