2023_InfraRed_Image_Restoration

This repository contains the code used for the manuscript "Image restoration of degraded time-lapse microscopy data mediated by infrared-imaging."

Repository structure:

• Figures: contains the code used to generate the figure of the paper.
• IR2_training_prediction: contains the code used to apply IR2 to microscopy images.
• Sample: example data.

Installation

Simply download the whole repository.

Requirements

The code has been tested with Python 3.9 and requires the following packages:

• numpy
• scikit-image
• tifffile
• matplotlib
• scipy
• pandas
• seaborn
• csbdeep

To install CSBDEEP with GPU, please follow the CSBDEEP documentation.

Using the repository

To run IR2 on your own data, you can use the scripts in the folder IR2_training_prediction in sequential order. Here we provide a step-by-step guide:

Data Management

Image data should be tif files accompanied by a txt metadata file. Data structure examples are provided in the Sample subfolders. Images should be arranged as follows:

__ sample_folder
 |
 |__ image_channel=ch00.tif
 |__ image_channel=ch01.tif
 |__ image_channel=ch[CCC]_params.txt

Where image_ch0.tif, image_ch1.tif are the visible and near infrared light images, respectively, and image_channel=ch[CCC]_params.txt contains at least the XYZ dimension of the tif dataset in the following format:

555:ROIWidth
555:ROIHeight
110:Planes

01_transfer_and_fuse_patch.py

This script extract patches of defined dimensions and creates a Dataset folder containing the images that will be used to train the IR2 network.

Parameters:

• create_Patches: boolean variable, default=True.
• inpaths: list of paths pointing to the raw data to process.
• patchSize: dimension of each patch. Default: (16,64,64).
• N_patches: number of patches to extract. Default: 100000.
• probmeths: probability distribution used to extract patch locations. Default: 'otsu', otherwise 'flat'.
• chOtsu: channel number to use to compute otsu threshold. Default: -1 (the near infrared).
• bias: if probmeths='otsu', defines the fraction of patches centered in bright regions of the sample. Default: 0.75.
• optimizeCoverage: whether to optimize sample coverage. Default: False.
• cThr: fraction of bright pixels to be covered by the patches. Default: 75%.
• nCoverage: number of times the bright pixels have to be covered by the patches. Default: 1.
• thresholdCorrelation: whether to filter out patches that show low correlation (e.g. misaligned or noise). Default: False.
• rThr: threshold of correlation. Default: 0.75.
• localRegister: whether to register the patches after extraction using correlation. Default: True.
• maskFilter: whether to filter out patches using a binary mask. Default: False.
• mask: binary mask. Default: None.

This script will generate a subfolder containing all patches extracted in tif and npz format, as well as the Otsu_mask generated:

__ sample_folder
 |
 |__DataSet
   |
   |__ image_channel=[CCC].npz
   |__ Otsu_mask_image_channel=[CCC].tif
   |__ tif_gt
   | |
   | |__ patch_001.tif
   | |__ patch_002.tif
   | |__ ...
   |
   |__tif_input
     |
     |__ patch_001.tif
     |__ patch_002.tif
     |__ ...

02_train_net_csbdeep.py

Script to train a standard UNet CARE network.

Parameters:

• pathsData: list of paths pointing to the datasets used for training.
• pathModel: path to CARE model.
• modelName
• N_max: set maximum number of patches. Default: None.
• train_batch_size: Default: 8.
• unet_n_depth: number of UNet layers. Default: 2.

03_predict_patches.py

Use the trained model to predict the patches for a list of datasets. Can use multiple models to compare results.

Parameters:

• pathsData
• pathModels
• modelNames

04_predict_full.py

Use the trained model to predict the whole tif images for a list of datasets. Can use multiple models to compare results.

Parameters:

• pathsData
• pathModels
• modelNames

05_quantify_patches.py

For each dataset, generates a csv file containing the following information for each patch reconstructed:

• Information content (IC)
• Structural similarity index (SSMI, relative to ground truth)
• Root mean square error (RMSE, relative to ground truth)

06_restore_TL.py

Restore timelapse datasets, which should be arranged as the training dataset, except that only one channel is provided, and one parameter file is provided per timepoint:

__ sample_folder
 |
 |__ image,_t=001,_channel=ch00,.tif
 |__ image,_t=001,_channel=ch[CCC]_params.txt
 |__ image,_t=002,_channel=ch00,.tif
 |__ image,_t=002,_channel=ch[CCC]_params.txt
 |__ ...