DenoisET implements the Noise2Noise algorithm [1] for cryoET data. While this machine learning algorithm has previously been implemented in software packages such as Cryo-CARE [2] and Topaz-Denoise [3], DenoisET is designed to streamline integration with AreTomo3 and provide a more automated approach that algorithmically decides which tomograms to select for training and when to transition from training to inference. DenoisET can be run during live data collection, either by applying a previously-trained model for inference or training a new 3D denoising model from scratch. We recommend using the CTF-deconvolved tomograms reconstructed by AreTomo3 for both training and inference. More details are available in Ref. [4].
The following will clone DenoisET and install it in a new conda environment.
git clone https://github.com/apeck12/denoiset.git
cd denoiset
conda create --name denoiset python=3.11.4
conda activate denoiset
pip install .
The following command runs both training and inference, with training tomograms selected based on their quality metrics and the transition to inference automatically determined to avoid checkerboard and other overdenoising artifacts:
denoise3d --input [tomogram directory] --output [output directory] --metrics_file [path to TiltSeries_Metrics.csv]
By default, the file nomenclature is expected to follow AreTomo3's conventions, with paired half tomograms ending in ODD_Vol.mrc and EVN_Vol.mrc. Only tomograms without EVN or ODD in their filenames will be denoised during inference. All tomograms (even, odd and full) are expected to be in the same input directory. A training subdirectory will be created in the output directory and populated with the list of training tomograms, the per-epoch model files (epoch[n].pth), per-epoch statistics (training_stats.csv), and representative denoised tomograms during each training epoch.
To adjust the volume of training data, the --n_extract (default: 250) and --max_selected (default: 30) parameters respectively determine the number of subvolumes extracted per tomogram per training epoch and the maximum number of tomograms to use during training. The default minimum number of tomograms (--min_selected) is 10. The default number of training epochs (--n_epochs) is 20, but training will automatically be terminated and inference started once the threshold set by --ch_threshold (default 0.034) is reached.
To run this concurrently with AreTomo3 processing, the --live flag can be supplied. The --t_interval and --t_exit flags determine how frequently DenoisET checks for new tomograms and how long it waits to exit after finding none. The default values for these are 5 and 30 minutes, respectively, but the parameters should be given in seconds.
The following flags can be used to adjust the default thresholds for the quality metrics:
--tilt_axis, maximum deviation from the median value in degrees (default: 1.5)--thickness, minimal sample thickness in Angstrom (default: 1000)--global_shift, maximum global shift in Angstrom (default: 750)--bad_patch_low, maximum fraction of bad patches at low tilt angles (default: 0.05)--bad_patch_all, maximum fraction of bad patches across the tilt range (default: 0.1)--ctf_res, maximum resolution of the CTF score in Angstrom for the zero-tilt image (default: 20)--ctf_score, minimum CTF score for the zero-tilt image (default: 0.05)
Running denoise3d --help will list additional parameters that can be adjusted from the command line.
The following pre-trained models that were trained on data collected at CZII can be found in the models directory. With the exception of the cilia model, all models were trained on 5 Å CTF-deconvolved tomograms.
- cilia.pth: data from mouse olfactory neuronal cilia
- lysosome.pth: data are from affinity-purified lysosomes
- minicell.pth: data are available on the CryoET Data Portal under deposition ID CZCDP-10312
- synaptosome.pth: data are available on the CryoET Data Portal under deposition ID CZCDP-10313
- phantom.pth: data are available on the CryoET Data Portal under deposition ID CZCDP-10310
[1] Lehtinen, J. et al. (2018) Noise2Noise: Learning Image Restoration without Clean Data. arXiv: 1803.04189.
[2] Buccholz, T., Jordan, M., Pigino, G. and Jug, F. (2018) Cryo-CARE: Content-Aware Image Restoration for Cryo-Transmission Electron Microscopy Data. arXiv:1810.05420.
[3] Bepler, T., Kelley, K., Noble, A. J, and Berger, B. (2020) Topaz-Denoise: general deep denoising models for cryoEM and cryoET. Nature Communications: 11, 5208.
[4] Peck, A. et al. (2025) AreTomoLive: Automated reconstruction of comprehensively-corrected and denoised cryo-electron tomograms in real-time and at high throughput. (in preparation)