Tyger

Appearance

Gadgetron examples ​

Gadgetron is an open source project for medical image reconstruction and can be used with Tyger. These examples assume that you have ismrmrd installed.

Basic example ​

Generate some test data with:

bash
ismrmrd_generate_cartesian_shepp_logan

This will generate a testdata.h5 file. You can then run a reconstruction with:

bash
ismrmrd_hdf5_to_stream -i testdata.h5 --use-stdout \
| tyger run exec -f basic_gadgetron.yml \
| ismrmrd_stream_to_hdf5 --use-stdin -o out_basic.h5

basic_gadgetron.yml looks like this:

yaml
job:
  codespec:
    image: ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda:latest
    buffers:
      inputs:
        - input
      outputs:
        - output
    args:
      - "-c"
      - "default.xml"
      - "--from_stream"
      - "-i"
      - "$(INPUT_PIPE)"
      - "-o"
      - "$(OUTPUT_PIPE)"
    resources:
      requests:
        cpu: 1

Using dependent measurements ​

It is common in MRI to perform a noise reference scan before the main scan. The noise scan is used to estimate a noise covariance matrix and the subsequent scan is reconstructed after noise pre-whitening using this covariance matrix.

For this example, download the noise and main scan raw data files:

bash
curl -OL https://aka.ms/tyger/docs/samples/dependencies/noise_scan.h5
curl -OL https://aka.ms/tyger/docs/samples/dependencies/main_scan.h5

Compute noise covariance matrix ​

Start by creating buffers for the noise data and covariance matrix:

bash
noise_buffer_id=$(tyger buffer create)
noise_covariance_buffer_id=$(tyger buffer create)

Then write the noise data to the buffer:

bash
ismrmrd_hdf5_to_stream -i noise_scan.h5 --use-stdout \
    | tyger buffer write $noise_buffer_id

Now compute the noise covariance matrix using the two buffers:

bash
tyger run exec -f noise_gadgetron.yml \
    -b input=$noise_buffer_id \
    -b noisecovariance=$noise_covariance_buffer_id

noise_gadgetron.yml looks like this:

yaml
job:
  codespec:
    image: ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda:latest
    buffers:
      inputs:
        - input
      outputs:
        - noisecovariance
    args:
      - "-c"
      - "default_measurement_dependencies.xml"
      - "--from_stream"
      - "-i"
      - "$(INPUT_PIPE)"
      - "-o"
      - /dev/null
      - --disable_storage
      - 'true'
      - --parameter
      - noisecovarianceout=$(NOISECOVARIANCE_PIPE)

Reconstruction ​

We are ready to reconstruct the main scan data by referencing the noise covariance matrix buffer:

bash
ismrmrd_hdf5_to_stream -i main_scan.h5 --use-stdout \
    | tyger run exec -f snr_gadgetron.yml --logs -b noisecovariance=$noise_covariance_buffer_id  \
    | ismrmrd_stream_to_hdf5 --use-stdin -o main-scan-recon.h5

snr_gadgetron.yml looks like this (note the two input buffers):

yaml
job:
  codespec:
    image: ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda:latest
    buffers:
      inputs:
        - input
        - noisecovariance
      outputs:
        - output
    args:
      - "-c"
      - "Generic_Cartesian_FFT.xml"
      - "--from_stream"
      - "-i"
      - "$(INPUT_PIPE)"
      - "-o"
      - "$(OUTPUT_PIPE)"
      - --disable_storage
      - 'true'
      - --parameter
      - noisecovariancein=$(NOISECOVARIANCE_PIPE)

Distributed reconstruction ​

Next is an example that uses a distributed run. First, download the raw data:

bash
curl -OL https://aka.ms/tyger/docs/samples/binning.h5

Then run:

bash
ismrmrd_hdf5_to_stream -i binning.h5 --use-stdout \
| tyger run exec -f distributed_gadgetron.yml --logs \
| ismrmrd_stream_to_hdf5 --use-stdin -o out_binning.h5

distributed_gadgetron.yml looks like this:

yaml
job:
  codespec:
    image: ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda:latest
    buffers:
      inputs:
        - input
      outputs:
        - output
    args:
      - "-c"
      - "CMR_2DT_RTCine_KspaceBinning_Cloud.xml"
      - "--from_stream"
      - "-i"
      - "$(INPUT_PIPE)"
      - "-o"
      - "$(OUTPUT_PIPE)"
    env:
      GADGETRON_REMOTE_WORKER_COMMAND: printenv TYGER_GADGETRON_WORKER_ENDPOINT_ADDRESSES
    resources:
      requests:
        cpu: 1

worker:
  codespec:
    image: ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda:latest
    args: []
    endpoints:
      gadgetron: 9002
    resources:
      requests:
        cpu: 3000m
        memory: 4G
      limits:
        memory: 4G
  replicas: 2

This run is made up of a job with two worker replicas.