Skip to main content

cryoCARE is a deep learning approach for cryo-TEM tomogram denoising.

Project description


This package is a memory efficient implementation of cryoCARE. This setup trains a denoising U-Net for tomographic reconstruction according to the Noise2Noise training paradigm. Therefor the user has to provide two tomograms of the same sample. The simplest way to achieve this is with direct-detector movie-frames. These movie-frames can be split in two halves (e.g. with MotionCor2 -SplitSum 1 or with IMOD alignframes -debug 10000) from which two identical, up to random noise, tomograms can be reconsturcted. These two (even and odd) tomograms can be used as input to this cryoCARE implementation.


cryoCARE uses .json configuration files and is run in three steps:

1. Prepare Training Data

To prepare the training data we have to provide all tomograms on which we want to train. Create an empty file called train_data_config.json, copy-paste the following template and fill it in.

  "even": [
  "odd": [
  "patch_shape": [
  "num_slices": 1200,
  "split": 0.9,
  "tilt_axis": "Y",
  "n_normalization_samples": 500,
  "path": "./"


  • "even": List of all even tomograms.
  • "odd": List of all odd tomograms. Note the order has to be the same as in "even".
  • "patch_shape": Size of the sub-volumes used for training. Should not be smaller than 64, 64, 64.
  • "num_slices": Number of sub-volumes extracted per tomograms.
  • "tilt_axis": Tilt-axis of the tomograms. We split the tomogram along this axis to extract train- and validation data separately.
  • "n_normalization_samples": Number of sub-volumes extracted per tomograms, which are used to compute mean and standard deviation for normalization.
  • "path": Path to which the training and validation data is saved.

Run Training Data Preparation:

After installation of the package we have access to built in Python-scripts which we can call. To run the training data preparation we run the following command: --conf train_data_config.json

2. Training

Create an empty file called train_config.json, copy-paste the following template and fill it in.

  "train_data": "./",
  "epochs": 100,
  "steps_per_epoch": 200,
  "batch_size": 16,
  "unet_kern_size": 3,
  "unet_n_depth": 3,
  "unet_n_first": 16,
  "learning_rate": 0.0004,
  "model_name": "model_name",
  "path": "./"


  • "train_data": Path to the directory containing the train- and validation data. This should be the same as the "path" from above.
  • "epochs": Number of epochs used to train the network.
  • "steps_per_epoch": Number of gradient steps performed per epoch.
  • "batch_size": Used training batch size.
  • "unet_kern_size": Convolution kernel size of the U-Net. Has to be an odd number.
  • "unet_n_depth": Depth of the U-Net.
  • "unet_n_first": Number of initial feature channels.
  • "learning_rate": Learning rate of the model training.
  • "model_name": Name of the model.
  • "path": Path to which the model is saved.

Run Training:

To run the training we run the following command: --conf train_config.json

3. Prediction

Create an empty file called predict_config.json, copy-paste the following template and fill it in.

  "model_name": "model_name",
  "path": "./",
  "even": "/path/to/even.rec",
  "odd": "/path/to/odd.rec",
  "n_tiles": [1, 1, 1],
  "output_name": "denoised.rec"

Note: Currently only a single tomogram can be denoised at a time i.e. if you want to denoise multiple tomograms you have to run this step for each tomo individually.


  • "model_name": Name of the trained model which should be used.
  • "path": Path to the parent directory where the model is stored. This corresponds to "path" in the train_config.json.
  • "even": Path to the even tomogram.
  • "odd": Path to the odd tomogram.
  • "n_tiles": Initial tiles per dimension. Gets increased if the tiles do not fit on the GPU.
  • "output_name": Name of the denoised tomogram.

Run Prediction:

To run the training we run the following command: --conf predict_config.json


Please first install TensorFlow 2 by following the official instructions.

Then you can install cryoCARE with pip: pip install cryoCARE

Note: I would recommend to use miniconda or singularity to manage environments and versions.

How to Cite

  title={Cryo-CARE: content-aware image restoration for cryo-transmission electron microscopy data},
  author={Buchholz, Tim-Oliver and Jordan, Mareike and Pigino, Gaia and Jug, Florian},
  booktitle={2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)},

  title={Content-aware image restoration for electron microscopy.},
  author={Buchholz, Tim-Oliver and Krull, Alexander and Shahidi, R{\'e}za and Pigino, Gaia and J{\'e}kely, G{\'a}sp{\'a}r and Jug, Florian},
  journal={Methods in cell biology},

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for cryoCARE, version 0.1.1
Filename, size File type Python version Upload date Hashes
Filename, size cryoCARE-0.1.1-py3-none-any.whl (16.7 kB) File type Wheel Python version py3 Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page