Table of Contents

• 1. Getting Started
  • 1.1. Reconstruct a phantom raw data set using the MRD client/server pair
  • 1.2. Creating a custom reconstruction/analysis module
    • 1.2.1. Adding a raw k-space filter
    • 1.2.2. Adding an image processing filter
  • 1.3. Using raw data from an MRI scanner
  • 1.4. Using DICOM images as input data
• 2. Setting up a working environment for the Python MRD client/server
  • 2.1. Setting up a devcontainer environment
  • 2.2. Setting up a conda environment
  • 2.3. Setting up a Docker environment
• 3. Code design
• 4. Saving incoming data
• 5. Startup scripts

1. Getting Started

1.1. Reconstruct a phantom raw data set using the MRD client/server pair

1. Set up a working environment for the MRD client/server pair using conda or Docker.

2. In a command prompt, generate a sample raw dataset:

   python generate_cartesian_shepp_logan_dataset.py -o phantom_raw.h5
   

   For a Docker environment, this should be performed in the same Docker container as the client.

   MRD data is stored in the HDF file format in a hierarchical structure in groups. The above example creates a dataset group containing:

   /dataset/data   Raw k-space data
   /dataset/xml    MRD header
   

3. Start the server in verbose mode by running:

   python main.py -v
   

   For a Docker environment, this should be performed in a separate Docker container.

4. Start the client and send data to the server for reconstruction:

   python client.py -G dataset -o phantom_img.h5 phantom_raw.h5
   

   The -G argument specifies the group name in the output file, the -o argument specifies the output file, and the last argument is the input file.

   If using conda, this should be run in a new command prompt. For Docker, this can be run in the same container from step 1.

   MRD image data are also stored in HDF files arranged by groups. If the -G argument is not provided, a group name will be automatically created with the current date and time. This may be useful when running the client multiple times, as multiple groups, organized by date/time, can be stored in the same output file. Images are further grouped by series index, with a sub-group named image_x, where x is image_series_index in the ImageHeader. For example:

   /dataset/image_0/data         Image data
   /dataset/image_0/header       MRD ImageHeader structure
   /dataset/image_0/attributes   MRD MetaAttributes text
   

5. The mrd2gif.py program can be used to convert an MRD Image file into an animated GIF for quick previewing:

   python mrd2gif.py phantom_img.h5
   

   A GIF file (animated if multiple images present) is generated in the same folder as the MRD file with the same base file name and the group and sub-groups appended.

   The reconstructed images in /tmp/phantom_raw.h5 can be opened in any HDF viewer such as https://www.hdfgroup.org/downloads/hdfview/. In Python, the ismrmrd-python-tools repository has an interactive imageviewer tool that can be used to view MRD formatted HDF files. The syntax is:

   python imageviewer.py phantom_img.h5
   

   In MATLAB, the ISMRMRD library contains helper classes to load and view MRD files. The files can also be read using MATLAB’s built-in HDF functions:

   img = h5read('/tmp/phantom_img.h5', '/dataset/image_0/data');
   figure, imagesc(img), axis image, colormap(gray)
   

1.2. Creating a custom reconstruction/analysis module

The MRD server has a modular design to allow for easy integration of custom reconstruction or image analysis code. The config that is passed by the client (e.g. the --config (-c) argument in client.py) is interpreted by the server as the "module" that should be executed to parse the incoming data. For example, a config of invertcontrast will select invertcontrast.py as the module to be run. Additional modules can be added simply be creating the appropriately named .py file in the Python path (e.g. the current folder).

If a file/module corresponding to the selecting config cannot be found, the server will fall back to a default config. The default config can be provided to main.py using the --defaultConfig (-d) argument. It is recommended that the default config argument be set in the CMD line of the Dockerfile when building an image to indicate the intended config to be run.

1.2.1. Adding a raw k-space filter

In this example, a Hanning filter is applied to raw k-space data.

1. Create a copy of invertcontrast.py named filterkspace.py. For other workflows, it may be preferable to start with another example such as simplefft.py.

2. The NumPy library contains a Hanning filter that can be applied to k-space data before the Fourier transform.

3. In the process_raw() function the filterkspace.py file, find the section where raw k-space data is already sorted into a Cartesian grid, just prior to Fourier transform. Add a Hanning filter and perform element-wise multiplication of the k-space data with the filter:

   Old code:

   # Fourier Transform
   data = fft.fftshift( data, axes=(1, 2))
   data = fft.ifft2(    data, axes=(1, 2))
   data = fft.ifftshift(data, axes=(1, 2))
   

   New code:

   # Apply Hanning filter
   logging.info("Applying Hanning filter to k-space data")
   filt = np.sqrt(np.outer(np.hanning(data.shape[1]), np.hanning(data.shape[2])))
   filt = np.expand_dims(filt, axis=(0,3))
   data = np.multiply(data,filt)
   np.save(debugFolder + "/" + "rawFilt.npy", data)
   
   # Fourier Transform
   data = fft.fftshift( data, axes=(1, 2))
   data = fft.ifft2(    data, axes=(1, 2))
   data = fft.ifftshift(data, axes=(1, 2))
   

4. The module used by the server is specified by the config option on the client side. The Server class used in this code attempts to find a Python module matching the name of the config file if it doesn't match one of the default examples. Start the server and in a separate window, run the client with the -c filterkspace option to specify the new config:

   python client.py -c filterkspace -o phantom_img.h5 phantom_raw.h5
   

5. Create a GIF preview of the filtered image:

   python mrd2gif.py phantom_img.h5
   

1.2.2. Adding an image processing filter

In this example, a high-pass filter is applied to images.

1. Create a copy of invertcontrast.py named filterimage.py. For other workflows, it may be preferable to start with another example such as analyzeflow.py.

2. The Python Pillow library contains a high-pass filter named FIND_EDGES. Add the following line to the top of the newly created filterimage.py to import this library:

   from PIL import Image, ImageFilter
   

3. In the process_image() function the filterimage.py file, find the sections where the incoming images are being normalized and filtered. The Pillow image filters require images with values in the range 0-255, so replace these two sections as follows:

   Old code:

   # Normalize and convert to int16
   data = data.astype(np.float64)
   data *= maxVal/data.max()
   data = np.around(data)
   data = data.astype(np.int16)
   
   # Invert image contrast
   data = maxVal-data
   data = np.abs(data)
   data = data.astype(np.int16)
   np.save(debugFolder + "/" + "imgInverted.npy", data)
   

   New code:

   # Normalize to range 0-255
   data = data.astype(np.float64)
   data *= 255/data.max()
   
   # Apply a 2D high-pass filter for each image
   logging.info("Applying high-pass image filter")
   from PIL import Image, ImageFilter
   for iImg in range(data.shape[-1]):
       im = Image.fromarray(np.squeeze(data[...,iImg])).convert('RGB')
       im = im.filter(ImageFilter.FIND_EDGES)
       data[:,:,0,0,iImg] = np.asarray(im)[...,0]
   
   # Rescale back to 16-bit
   data = data * maxVal/data.max()
   data = data.astype(np.int16)
   np.save(debugFolder + "/" + "imgFiltered.npy", data)
   

4. The module used by the server is specified by the config option on the client side. The Server class used in this code attempts to find a Python module matching the name of the config file if it doesn't match one of the default examples. Start the server and in a separate window, run the client with the -c filterimage option to specify the new config:

   python client.py -c filterimage -o phantom_img.h5 phantom_raw.h5
   

5. Create a GIF preview of the filtered image:

   python mrd2gif.py phantom_img.h5
   

1.3. Using raw data from an MRI scanner

Raw data from MRI scanners can be converted into MRD format using publicly available conversion tools such as siemens_to_ismrmrd, ge_to_ismrmrd, philips_to_ismrmrd, and bruker_to_ismrmrd. These can be used as input data for the client as part of streaming MRD framework.

For Siemens data, raw data in .dat format can be converted and processed as follows:

1. Convert a raw data file named gre.dat into MRD format:

   siemens_to_ismrmrd -Z -f gre.dat -o gre_raw.h5
   

   The following options are used:

   -Z  Convert all acquisitions from a multi-measurement (multi-RAID) file
   -f  Input file (Siemens raw data .dat file)
   -o  Output file (MRD raw data .h5 file)
   

   If the input file is a multi-RAID file, then several output files are created such as gre_raw_1.h5 and gre_raw_2.h5. The first measurements are dependency data while the main acquisition is in the last numbered file.

2. Start the server and in a separate window, run the client using the converted file:

   python client.py -c invertcontrast -o gre_img.h5 gre_raw_2.h5
   

   Note that the invertcontrast example module only does basic Fourier transform reconstruction and does not support undersampling or more complex acquisitions.

1.4. Using DICOM images as input data

For image processing workflows, DICOM images can be used as input by converting them into MRD format. The dicom2mrd.py script can be used to convert DICOMs to MRD, while mrd2dicom.py can be used to perform the inverse.

1. Create a folder containing DICOM files with file extensions .ima or .dcm. Files can also be organized in sub-folders if desired.

2. Run the dicom2mrd conversion script:

   python dicom2mrd.py -o dicom_img.h5 dicoms
   

   Where the DICOM files are in a folder called dicoms and an output file dicom_img.h5 is created containing the MRD formatted images.

3. Start the server and in a separate window, run the client using the converted file:

   python client.py -c invertcontrast -o dicom_img_inverted.h5 dicom_img.h5
   

4. Convert the output MRD file back to a folder of DICOMs:

   python mrd2dicom.py dicom_img_inverted.h5
   

2. Setting up a working environment for the Python MRD client/server

2.1. Setting up a devcontainer environment

Development containers (devcontainers) are a convenient way of using a Docker image as a working environment instead of installing packages locally. If the Dev Containers extension is installed in Visual Studio Code, then a prompt will appear when opening this folder to re-open it in a devcontainer. The devcontainer is also compatible with GitHub Codespaces. Further details can be found in doc/devcontainers.md.

2.2. Setting up a conda environment

Conda is a Python environment manager that is useful for creating and maintaining Python packages and their dependencies. It is available either as part of the larger Anaconda product, or separately as part of Miniconda. Although not required, it's helpful in setting up an environment for the Python ISMRMD client/server. Mamba and micromamba are drop-in replacements for Conda that are often faster at resolving dependencies. The following instructions are for micromamba, but the command micromamba can be replaced with conda if conda is preferred.

1. Download and install Git for your operating system.

2. Download and install micromamba or Miniconda for your operating system.

3. For Windows, open a powershell prompt. In MacOS and Linux, open a standard command prompt.

4. Clone (download) this repository

   git clone https://github.com/kspaceKelvin/python-ismrmrd-server.git
   

5. Change into this respository's directory and create a new conda environment for MRD using the dependencies listed in environment.yml. If using Windows, then environment_windows.yml should be used instead.

   cd python-ismrmrd-server
   micromamba create -f environment.yml
   

6. Active the new MRD environment

   micromamba activate mrd
   

7. If using Windows, then install the ISMRMRD Python library through pip. For MacOS and Linux, this was installed through conda.

   pip install ismrmrd
   

To use this environment in the future, open a command prompt and run micromamba activate mrd.

2.3. Setting up a Docker environment

Docker is a virtualization platform that allows software to run in isolated environments called containers. It provides a convenient mechanism to package up a reconstruction program and all its libraries in a manner that can be easily deployed to other computers without manually installing dependencies or other configuration steps.

A complete working environment of this respository has been compiled into a Docker image stored on Docker Hub. This can be used to quickly get started, but setting up a native Python environment is recommended for development work.

1. Download and install Docker with the standard settings.

2. Download the Docker image of this repository by opening a command prompt and running:

   docker pull kspacekelvin/fire-python
   

3. The Python MRD server can be started in a Docker container by running:

   In Windows:
       docker run -p=9002:9002 --rm -it -v C:\tmp:/tmp kspacekelvin/fire-python
   
   In MacOS/Linux:
       docker run -p=9002:9002 --rm -it -v /tmp:/tmp kspacekelvin/fire-python
   

   The command line options used are:

   -p=9002:9002      Allows access to port 9002 inside the container from port 9002 on the host.  
                     Change the first number to change the host port.
   -it               Enables "interactive" mode with a pseudo-tty.  This is necessary for "ctrl-c" to
                     stop the program.
   --rm              Remove the container after it is stopped
   -v C:\tmp:/tmp    Maps the C:\tmp folder on the host to /tmp inside the container.
                     Change the first path if using a different folder on the host computer.
                     Log and debug files are stored in this folder.
   

   The server can be stopped by pressing ctrl-c.

4. The Python MRD client can also be run in a Docker container (the internal IP address host.docker.internal needs to be manually added to the host in the latest version of Docker). If the server is running in a Docker container already, open a new command prompt and run:

   In Windows:

   docker run --rm -it --add-host=host.docker.internal:host-gateway -v C:\tmp:/tmp kspacekelvin/fire-python /bin/bash
   

   In MacOS/Linux:

   docker run --rm -it --add-host=host.docker.internal:host-gateway -v /tmp:/tmp kspacekelvin/fire-python /bin/bash
   

   In this invocation, the /bin/bash argument is used to start the container with a bash shell prompt instead of starting the Python MRD server. The client can be called by running:

   python /opt/code/python-ismrmrd-server/client.py -a host.docker.internal -p 9002 -o /tmp/phantom_img.h5 /tmp/phantom_raw.h5
   

   The command line options used are:

   -a host.docker.internal  Send data to address host.docker.internal, which
                            resolves to the IP address of the Docker host.
   -p 9002                  Send data to port 9002.  This is sent to the host, which
                            is then redirected to the server container due to the
                            "-p" port mapping when starting the server. 
   -o                       Specifies the output file name
   -G                       Specifies the group name in the output file
   

   This Docker container can also be used to run the generate_cartesian_shepp_logan_dataset.py script to generate phantom data:

       python /opt/code/python-ismrmrd-server/generate_cartesian_shepp_logan_dataset.py -o /tmp/phantom_raw.h5
   

3. Code design

This code is designed to provide a reference implementation of an MRD client/server pair. It is modular and can be easily extended to include additional reconstruction/analysis programs.

• main.py: This is the main program, parsing input arguments and starting a "Server" class instance.

• server.py: The "Server" class determines which reconstruction algorithm (e.g. "simplefft" or "invertcontrast") is used for the incoming data, based on the requested config information.

• connection.py: The "Connection" class handles network communications to/from the client, parsing streaming messages of different types, as detailed in the MRD documentation. The connection class also saves incoming data to MRD files if this option is selected.

• constants.py: This file contains constants that define the message types of the MRD streaming data format.

• mrdhelper.py: This class contains helper functions for commonly used MRD tasks such as copying header information from raw data to image data and working with image metadata.

• client.py: This script can be used to function as the client for an MRD streaming session, sending data from a file to a server and saving the received images to a different file. Additional description of its usage is provided below.

• generate_cartesian_shepp_logan_dataset.py: Creates an MRD raw data file of a Shepp-Logan phantom with Cartesian sampling. Borrowed from the ismrmrd-python-tools repository.

• dicom2mrd.py: This program converts a folder of DICOM images to an MRD image .h5 file, allowing DICOM images to be use as input for MRD streaming data.

• mrd2dicom.py: This program converts an MRD image .h5 file to a folder of DICOM images.

• mrd2gif.py: This program converts an MRD image .h5 file into an animated GIF for quick previews.

There are several example "modules" that can be selected by specifying their name via the config (-c) argument:

• invertcontrast.py: This program accepts both incoming raw data as well as image data. The image contrast is inverted and images are sent back to the client.

• simplefft.py: This file contains code for performing a rudimentary image reconstruction from raw data, consisting of a Fourier transform, sum-of-squares coil combination, signal intensity normalization, and removal of phase oversampling.

• rgb.py: This program accepts incoming image data, applies a jet colormap, and sends RGB images back to the client.

• analyzeflow.py: This program accepts velocity phase contrast image data and performs basic masking.

4. Saving incoming data

It may be desirable for the MRD server to save a copy of incoming data from the client. For example, if the client is an MRI scanner, then the saved data can be used for offline simulations at a later time. This may be particularly useful when the MRI scanner client is sending image data, as images are not stored in a scanner's raw data file and would otherwise require offline simulation of the MRI scanner reconstruction as well.

The feature may be turned on by starting the server with the -s option (disabled by default). Data files are named by the current date/time and stored in /tmp/share/saved_data. The saved data folder can be changed using the -S option. For example, to turn on saving of incoming data in the /tmp folder, start the server with:

python main.py -s -S /tmp

The -s flag is enabled in the startup script start-fire-python-server-with-data-storage.sh.

Alternatively, this feature can be enabled on a per-session basis when the client calls the server with the config savedataonly. In this mode, incoming data (raw or image) is saved, but no processing is done and no images are sent back to the client.

The resulting saved data files are in MRD .h5 format and can be used as input for client.py as detailed above.

5. Startup scripts

There are three scripts that may be useful when starting the Python server in a chroot environment (i.e. on the scanner). When using this server with FIRE, the startup script is specified in the fire.ini file as chroot_command. The scripts are:

• start-fire-python-server.sh: This script takes one optional argument, which is the location of a log file. If not provided, logging outputs are discarded.

• sync-code-and-start-fire-python-server.sh: This script copies all files from /tmp/share/code/ to /opt/code/python-ismrmrd-server/. These paths are relative to the chroot container and when run with FIRE, /tmp/share/code/ is a shared folder from the host computer at %CustomerIceProgs%\fire\share\code\. This "sync" step allows Python code to be modified on the host computer and executed by the Python reconstruction process. The Python reconstruction program is then started in the same way as in start-fire-python-server.sh. This is the startup script specified in the default fire.ini configuration file. However, this script should not be used in stable projects, as overwriting existing files with those in /tmp/share/code is likely undesirable.

• start-fire-python-server-with-data-storage.sh: This script is the same as start-fire-python-server.sh, but saves incoming data (raw k-space readouts or images) to files in /tmp/share/saved_data, which is on the host computer at %CustomerIceProgs%\fire\share\saved_data. This is useful to save a copy of MRD formatted data (particularly for image-based workflows) for offline development, but this option should be used carefully, as raw data can be quite large and can overwhelm the limited hard drive space on the Windows host.