cryoCARE

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. Therefore the user has to provide two tomograms of the same sample. The simplest way to achieve this is with direct-detector movie-frames.

You can use Warp to generate two reconstructed tomograms based on the even/odd frames. Alternatively, the 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 reconstructed.

These two (even and odd) tomograms can be used as input to this cryoCARE implementation.

Changelog

Version 0.3

• cyroCARE_train now supports parallelization over multiple GPUs.

Version 0.2

• cyroCARE_train produces a compressed and more portable model. This model can be copied and shared with others without relying on a certain folder structure.
• cryoCARE_predict supports to predict multiple tomograms in one run. Streamlined configuration with respect to the changes of cryoCARE_train.
• Streamlined installation instructions
• CUDA 11 support
• Minor changes/ fixed couple of bugs:
  • Proper padding of tomograms to avoid black frames in the denoised tomograms
  • Fix computation of validation cut off for small tomograms
  • Fix cryoCARE_predict if no tiling happens

Installation

Note: We assume that you have miniconda installed.

First you need to create a conda environment.

For CUDA 11:

conda create -n cryocare_11 python=3.8 cudatoolkit=11.0 cudnn=8.0 -c conda-forge
conda activate cryocare_11
pip install tensorflow==2.4
pip install cryoCARE

For CUDA 10:

conda create -n cryocare -c conda-forge -c anaconda python=3 keras-gpu=2.3.1
conda activate cryocare
pip install cryoCARE

Manual

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": [
    "/path/to/even.rec"
  ],
  "odd": [
    "/path/to/odd.rec"
  ],
  "mask": [
    "/path/to/mask.mrc"
  ],
  "patch_shape": [
    72,
    72,
    72
  ],
  "num_slices": 1200,
  "split": 0.9,
  "tilt_axis": "Y",
  "n_normalization_samples": 500,
  "path": "./"
}

Parameters:

• "even": List of all even tomograms.
• "odd": List of all odd tomograms. Note the order has to be the same as in "even".
• "mask": If desired, a list of binary masks to limit where subvolumes are extracted, similar to IsoNet. Can be left out to skip masking.
• "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": The training and validation data are saved here.

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: cryoCARE_extract_train_data.py --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": "./",
  "gpu_id": 0
}

Parameters:

• "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": Output path for the model.
• "gpu_id": This is optional. Provide the ID(s) of the GPUs you wish to use. Alternatively, you can specify the GPU ID(s) using the CUDA_VISIBLE_DEVICES environment variable. Training supports multiple GPUs (see below).

Run Training:

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

You will find a .tar.gz file in the directory you specified as path. This your model an will be used in the next step.

Train using multiple GPUs:

Training can be faster by running on multiple GPUs (which must be available in the same machine). Please note that the performance does not improve linearly with the number of devices used for training. The actual speedup will depend on your training settings and hardware.

There are two methods for specifying multiple GPU ID(s):

Method 1: Using gpu_id

You can specify multiple GPU ID(s) in the train_config.json file as follows (with 4 GPUs in this example):

"gpu_id": [0,1,2,3]

Note: This method takes precedence over the CUDA_VISIBLE_DEVICES method below. If you want to use that method, you should omit the "gpu_id" entry from your train_config.json file.

Method 2: Using the CUDA_VISIBLE_DEVICES environment variable

For example, with 4 GPUs:

export CUDA_VISIBLE_DEVICES=0,1,2,3
cryoCARE_train.py --conf train_config.json

Note: If running cryoCARE under a cluster resource manager such as SLURM, the CUDA_VISIBLE_DEVICES environment variable might be automatically set when you request a certain number of GPUs, so you don't need to set it explicitly. Check your cluster documentation or support team for details.

3. Prediction

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

{
  "path": "path/to/your/model/model_name.tar.gz",
  "even": "/path/to/even.rec",
  "odd": "/path/to/odd.rec",
  "n_tiles": [1,1,1],
  "output": "denoised.rec",
  "overwrite": false,
  "gpu_id": 0
}

Parameters:

• "path": Path to your model file.
• "even": Path to directory with even tomograms or a specific even tomogram or a list of specific even tomograms.
• "odd": Path to directory with odd tomograms or a specific odd tomogram or a list of specific odd tomograms in the same order as the even tomograms.
• "n_tiles": Initial tiles per dimension. Gets increased if the tiles do not fit on the GPU.
• "output": Path where the denoised tomograms will be written.
• "overwrite": Allow previous files to be overwritten.
• "gpu_id": This is optional. Provide the ID of the GPU you wish to use. Alternatively, you can specify the GPU ID using the CUDA_VISIBLE_DEVICES environment variable. Note that prediction only supports a single GPU currently.

Run Prediction:

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

How to Cite

@inproceedings{buchholz2019cryo,
  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)},
  pages={502--506},
  year={2019},
  organization={IEEE}
}

@article{buchholz2019content,
  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},
  volume={152},
  pages={277--289},
  year={2019}
}