Modern denoisers are typically trained on a representative dataset, sometimes consisting of just unpaired noisy shots. However, when data are acquired in real time to track dynamic cellular processes, it is not always practical nor economical to generate these training sets. Recently blind zero-shot denoisers have emerged that allow us to denoise single images without a training set or knowledge about the underlying noise. But such methods are currently too slow to be integrated into imaging pipelines that require rapid, real-time hardware feedback. Here we present Noise2Fast, which can overcome these limitations. Noise2Fast uses a novel downsampling technique we refer to as ‘chequerboard downsampling’. This allows us to train on a discrete 4-image training set, while convergence can be monitored using the original noisy image. Noise2Fast is faster than all similar methods with only a small drop in accuracy compared to the gold standard.

In other words Noise2Fast denoises single noisy images without any knowledge of the underlying noise distribution and without sensitive hyperparameters that need to be adjusted on a case-by-case basis (except learning rate in certain cases). It starts with randonly initialized weights, and everything it knows about the universe comes solely from the noisy image you feed it. Once it has denoised that image, weights get randomly initialized again and it starts the process anew on the next image.

If you don't already have anaconda, install it by following instructions at this link: https://docs.anaconda.com/anaconda/install/.

It would also be helpful to have ImageJ installed: https://imagej.nih.gov/ij/download.html.

Open Anaconda Prompt (or terminal if on Mac/Linux) and enter the following commands to create a new conda environment and install the required packages:

conda create --name N2F
conda activate N2F
conda install -c pytorch pytorch=1.8.0
conda install -c conda-forge tifffile=2019.7.26.2

If the installs don't work, removing the specific version may fix this.

Create a folder in the master directory (the directory that contains N2F.py) and put your noisy images into it. Then open anaconda prompt/terminal and run the following:

cd <masterdirectoryname>
conda activate N2F
python N2F.py <noisyfolder>

Replacing "masterdirectoryname" with the full path to the directory that contains N2F.py, and replacing "noisyfolder" with the name of the folder containing images you want denoised. Results will be saved to the directory '_N2F'. Issues may arise if using an image format that is not supported by the tifffile python package, to fix these issues you can open your images in ImageJ and re-save them as .tif (even if they were already .tif, this will convert them to ImageJ .tif).

To run on anything other than 2D grayscale images, use N2F_4D.py. This supports an arbitrary number of dimensions, as long as the last two dimensions are x and y. For example, here we use it on a 16x6x2x250x250 (tzcxy) image:

cd <masterdirectoryname>
conda activate N2F
python N2F_4D.py livecells

Output is in ImageJ format.

To run N2F on the noisy confocal images, open a terminal in the master directory and run:

cd <masterdirectoryname>
python N2F.py Confocal_gaussianpoisson

The denoised results will be in the directory 'Confocal_gaussianpoisson_N2F'.

To run N2F on our other datasets we first need to add synthetic gasussian noise. For example to test N2F on Set12 with sigma=25 gaussian noise, we would first:

cd <masterdirectoryname>
python add_gaussian_noise.py Set12 25

This will create the folder 'Set12_gaussian25' which we can now denoise:

python N2F.py Set12_gaussian25

Which returns the denoised results in a folder named 'Set12_gaussian25_N2F'.

To find the PSNR and SSIM between a folder containing denoised results and the corresponding folder containing known ground truths (e.g. Set12_gaussian25_N2F and Set12 if you followed above), we need to install one more conda package:

conda activate N2F
conda install -c anaconda scikit-image=0.17.2

Now we measure accuracy with the code:

cd <masterdirectoryname>
python compute_psnr_ssim.py Set12_gaussian25_N2F Set12

You can replace 'Set12' and 'Set12_gaussian25' with any pair of denoised/ground truth folders (order doesn't matter). Average PSNR and SSIM will be returned for the entire set.

We can run DIP, Noise2Self and Ne2Ne in the N2F environment:

conda activate N2F
python DIP.py Confocal_gaussianpoisson
python N2S.py Confocal_gaussianpoisson
python Ne2Ne.py Confocal_gaussianpoisson

Lequyer, J., Philip, R., Sharma, A. et al. A fast blind zero-shot denoiser. Nat Mach Intell (2022).