NIDDL - Neuro Imaging Denoising Via Deep Learning

Deep denoising pushes the limit of functional data acquisition by recovering high SNR calcium traces from low SNR videos acquired using low laser power or smaller exposure time. Thus deep denoising enables faster and longer volumetric recordings. For more details, please check our paper.

1. Download using git clone -b pytorch https://github.com/shiveshc/NIDDL.git
2. Much easier to install and use (checkout example.ipynb).
3. Denoise in Napari by running napari_niddl.py

If you find our work useful, please cite

Chaudhary, S., Moon, S. & Lu, H. Fast, efficient, and accurate neuro-imaging denoising via supervised deep learning. Nat Commun 13, 5165 (2022). https://doi.org/10.1038/s41467-022-32886-w

Denoise whole-brain videos

Denoise mechanosensory neurites

Recover high SNR calcium traces

Contents

1. Installation
2. Additional system requirements
3. Train on new dataset
4. Denoise calcium activity recordings

Installation

Installation steps tested for Linux and Python 3.9

1. Download pytorch version of NIDDL repository

Clone pytorch branch of repository using git clone -b pytorch https://github.com/shiveshc/NIDDL.git.

2. Setting up env and installing dependencies

Open command line terminal as administrator and navigate to cloned repository path using cd ~/NIDDL.

Next run following commands -

1. conda env create -f environment.yml
2. conda activate niddl-env

Installation should take ~10 minutes.

Troubleshooting

environment.yml has been tested on Linux. Installation on MacOS or Windows may throw errors. In this case you can try following -

1. Try using environment_macos.yml for macos.

Or

1. cd ~/NIDDL
2. conda create -n niddl-env python=3.9
3. conda activate niddl-env
4. pip install ipykernel
5. jupyter notebook
6. Run 1st cell in example.ipynb, manually install required libraries using pip install <lib>

3. Usage

Napari

Try denoising in Napari

1. cd ~/NIDDL
2. conda activate niddl-env
3. python napari_niddl.py
4. Denoise NIDDL widget should be visible in napari viewer
5. Set noisy_img_path image and corresponding high SNR gt_image_path (optional) in the widget.
6. Set DataType approprately (corresponding models will be used to denoise images)
7. Click Denoise

Notebook

1. Create jupyter kernel in niddl-env using jupyter notebook
2. Follow instructions in example.ipynb

Additional system requirements

Model training using GPUs is much faster and thus preferred. To be able to use GPUs, suitable NVIDIA drivers and CUDA libraries must be installed. Please make sure your CUDA versions are compatible with Pytorch 2.0.1.

Train on new dataset

1. Create jupyter kernel within niddl-env.

2. Follow instructions in example.ipynb.

3. Structure of training data folder should be organised as below -

   data
   ├───gt_imgs
   │   ├───img_1
   │   │       z_1.tif
   │   │       z_2.tif
   │   │       ...
   │   │
   │   ├───img_2
   │   │       z_1.tif
   │   │       z_2.tif
   │   │       ...
   │   │
   │   └───img_3
   │           z_1.tif
   │           z_2.tif
   │           ...
   │
   └───noisy_imgs
       ├───img_1
       │       z_1.tif
       │       z_2.tif
       │       ...
       │
       ├───img_2
       │       z_1.tif
       │       z_2.tif
       │       ...
       │
       └───img_3
               z_1.tif
               z_2.tif
               ...
   

Denoise calcium activity recordings

1. Structure of functional recording video datasets should be organised as below -

   vid1
   └───noisy_imgs
       ├───img_1
       │       z_1.tif
       │       z_2.tif
       │       ...
       │
       ├───img_2
       │       z_1.tif
       │       z_2.tif
       │       ...
       │
       ├───img_3
       │       z_1.tif
       │       z_2.tif
       │       ...
       ...
   

   Here img_1, img_2 etc. can correspond to individual time-points in videos.

2. Follow instruction in example.ipynb.