DELICATE: NucEnhancedApp Enhance the Nucleus Images in live-cell embryos via Effective Deep Learning Model

1. Programme Introduction

This is a programme for running the fluorescence images of cell nucleus enhancement, especially for C. elegans. For other animals, you need to re-train the DNN model with about 4 embryos Ground Truth.

2. Running Steps

Make sure you have installed python3 (https://www.python.org/downloads/) and conda (https://docs.anaconda.com/free/miniconda/index.html) in your computer linux/osx/windows.

Then run the following command one by one in your terminal/command line/power shell.

The first time

• conda create --name EmbNucEnhancementPYEnvironment20240329Ver python=3.9.19
• conda activate EmbNucEnhancementPYEnvironment20240329Ver
• pip install PyQt5
• pip install stardist
• pip install nibabel
• pip install tensorflow
• python NucEnhanceApp.py

Second and future

• conda activate EmbNucEnhancementPYEnvironment20240329Ver
• python NucEnhanceApp.py
• group the correct raw nucleus tif images folder !

Then you get the

Configure the parameters according to your images and choose recommended Zhaolab model

CAUTIONS!

• You need at least 16GiB physical or swap memory free to run our DEEP LEARNING enhancing model!! Close other memory-consuming programs!
• Make sure your images are cropped, which means there is only one embryo in the time series images.
• The 3D X Y Z size should be tuned to range 150-300, 250-400, 120-250, respectively, and keep the resolution correct.
• You may need to train your model with your own cropped data (1 embryo) or contact me (zelinli6-c(at)my.cityu.edu.hk). The training processes are provides as followed. I will help you in training your own model for your lab (specific protein or microscopic).

Close the terminal directly

3. Result comparison

The output folder looks like:

The programme would enhance all recognized nucleus images and allow starrynite and acetree to trace the whole live-cell lineage .

The enhanced performance would increase as the cell number increase (The messier the image, the better enhancement).

4. Training With Your Data

4.1 Prepare the groundtruth

Please configure file 0_tiff2d_to_niigz3d_for_nuc_gt.py to packed the 2D tif to 3D niigz according to the naming system. Now, you have transformed your images and CDfiles to a niigz groundtruth (Like file GroundTruthTrainingDataset3EmbDuLab.zip on https://doi.org/10.6084/m9.figshare.26778475.v1)

4.2 Prepare training data

Please read and run jupyter notebook file 1_nuc_data.ipynb step by step to ensure the groundtruth training dataset is ready (Like file GroundTruthTrainingDataset6EmbZhaoLab.zip on https://doi.org/10.6084/m9.figshare.26778475.v1). You can view the GT here.

4.3 Training

Run jupyter notebook 2_nuc_training.ipynb. The trained model is placed at ./Training/models/stardist_nuc.

4.4 Move your model for running

Create a new folder under ./static/models/ with the name you want. Move config.json, weights_best.h5, and thresholds.json in ./Training/models/stardist_nuc to your new created folder. Now you can select your model when running.

5. Issues/Notes

• Please make sure your computer has at least 16GiB memory.

• The progress bar reflect not real progress. Actually, it takes about 1 hour (good cpu) to enhance a 300 time point >550-cell embryo. Please consider the real progress in the terminal or check the output folder.

• 

• The programme is running with only CPU and need no GPU.