See the author's git repository here: https://github.com/AyushSomani001/VirtualStain_I2I

Virtual Labeling of Mitochondria in Living Cells using Correlative Imaging and Physics-guided Deep Learning
Ayush Somani, ARIF AHMED SEKH, IDA S. OPSTAD, ÅSA BIRNA BIRGISDOTTIR, TRULS MYRMEL, BALPREET SINGH AHLUWALIA, ALEXANDER HORSCH, KRISHNA AGARWAL, and DILIP K. PRASAD
In Biomedical Optics Express (BOE) 2022
[Paper] [BibTex]

This is the code for Virtual Labeling of Mitochondria in Living Cells using Correlative Imaging and Physics-guided Deep Learning. Figure (A) outlines the proposed method. The red coloured numbers represent the principal modules and their sequence, and Figure (B) conditional GAN of the employed architecture. The details of the architecture are presented in section 2.3.

This code in this repository can be used to run training and inference of our model on a single machine, and can be adapted for distributed training. This README will explain how to:

🔬 Data Availability

The data is made available through UiT public repository for large datasets, Dataverse: [To-be-Updated]

The TIFF frames extracted from the .dv mircoscopy file can then be converted into PNG format of dimension '1024 x 1024' and subsequently randomly cropped into '256 x 256' using the Data_Gen.py file in the assets folder.

💻 Citing the code

If you use this code, please cite our paper:

Ayush Somani, Arif Ahmed Sekh, Ida S. Opstad, Åsa Birna Birgisdottir, Truls Myrmel, Balpreet Singh Ahluwalia, Alexander Horsch, Krishna Agarwal, and Dilip K. Prasad, "Virtual labeling of mitochondria in living cells using correlative imaging and physics-guided deep learning," Biomed. Opt. Express 13, 5495-5516 (2022)

📑 BibTeX:

@article{somani2022virtual,
  title={Virtual Labeling of Mitochondria in Living Cells using Correlative Imaging and Physics-guided Deep Learning},
  author={Somani, Ayush and Sekh, Arif Ahmed and Opstad, Ida S and Birgisdottir, {\AA}sa Birna and Myrmel, Truls and Ahluwalia, Balpreet Singh and Agarwal, Krishna and Prasad, Dilip K and Horsch, Alexander},
  journal = {Biomed. Opt. Express},
  number = {10},
  pages = {5495--5516},
  publisher = {Optica Publishing Group},
  year={2022},
  volume = {13},
  month = {Oct},
  year = {2022},
  publisher = {Optica Publishing Group},
  url = {http://opg.optica.org/boe/abstract.cfm?URI=boe-13-10-5495},
  doi = {10.1364/BOE.464177},
}

Prerequisites

• Python 3
• PyTorch 1.0.1 or higher, with NVIDIA CUDA Support
• Tensorflow 1.10.0 or higher
• Other required python packages specified by requirements.txt. See the Installation.

Installation

Clone this repository:

git clone https://github.com/AyushSomani001/VirtualStain_I2I.git

Create a conda environment for VS, and install the requirements. This includes the required python packages from requirements.txt. Most of the required packages have been included in the built-in anaconda package:

conda create -n VS --file VirtualStain_I2I\requirements.txt
conda activate VS

OR Easily setup the Conda environment by cloning the miniconda3 environment Somani available for download here.

conda create --name VS --clone _assets\miniconda3\envs\Somani
conda activate VS

Note: You may replace VS with the choice of the new environment name. Somani is the name of the existing environment that you want to copy. To verify that the copy was made:

conda info --envs

Dataset preparation

To replicate the experiments, you need to prepare your dataset as the following. Taking the Mitochondria dataset as an example and creating just the training and test set split.

VirtualStain_I2I\Dataset
                  ├── train_A
                  │   └── images (Contains BrightField images)
                  ├── train_B
                  │   └── images (Contains Fluroscence counterpart images)
                  ├── test_A
                  │   └── images
                  └── test_B
                      └── images

Note: The VirtualStain_I2I\Dataset folder contains a sample subset of 200 image pairs, divided into a 70-30 training set and a test set ratio. This is not the exhaustive dataset that was compiled and utilized for the results section of the study.

Training and evaluation.

To train the model:

python train.py --name <model_name> --gpu_ids 0 --no_lsgan --lambda_feat 5 --lambda_feat_k 50 --dataroot './VirtualStain_I2I/Dataset'  --niter_decay 500

You can add other data arguments like: "--gaussian_loss" or "--dice_bce" present in the ```VirtualStain_I2I/options/train_options.py''' option to experiment with other G_GAN Loss Functions.

To Test the model:

python test.py --name <model_name> --dataroot './VirtualStain_I2I/Dataset' 

📌 Acknowledgments

This code extensively draws inspiration from pytorch-CycleGAN-and-pix2pix.

📧 Contact the Author

ayush[dot]somani[at]uit.no