Add Python fastMRI converter

[naegelejd]

We'll stage development of a fastMRI -> mrd converter in this branch. Leave the PR a draft.

[Copilot]

Pull request overview

Adds an initial (draft) Python CLI tool to convert fastMRI HDF5 datasets into MRD streams, plus a small CI/workflow adjustment to pin the Miniforge version for conda builds.

Changes:

• Introduce fastmri_to_mrd.py to convert fastMRI headers/k-space and reconstructed RSS images into MRD (BinaryMrdWriter).
• Add basic dataset presence checks and synthesize MRD noise acquisitions from k-space edge lines.
• Pin miniforge-version in the conda GitHub Actions workflow.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 7 comments.

File	Description
python/mrd/tools/fastmri_to_mrd.py	New fastMRI→MRD converter tool (header conversion + k-space/image export).
.github/workflows/mrd_conda.yml	Pins Miniforge version for more reproducible conda CI/builds.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

np.std(np.abs(...)) is computed multiple times here (including inside the log message). Compute it once and reuse the value to avoid repeated full-array passes over k-space data (can be expensive for large datasets).

Suggested change

	if np.std(np.abs(last_line.data)) < np.std(np.abs(first_line.data)):
	acq = last_line
	else:
	acq = first_line
	acq.head.flags = mrd.AcquisitionFlags.IS_NOISE_MEASUREMENT
	if np.std(np.abs(acq.data)) >= 1e-5:
	logger.warning(f"Expected to find noise in slice {slice}, but data has significant variation with std {np.std(np.abs(acq.data))}.")
	first_line_std = np.std(np.abs(first_line.data))
	last_line_std = np.std(np.abs(last_line.data))
	if last_line_std < first_line_std:
	acq = last_line
	acq_std = last_line_std
	else:
	acq = first_line
	acq_std = first_line_std
	acq.head.flags = mrd.AcquisitionFlags.IS_NOISE_MEASUREMENT
	if acq_std >= 1e-5:
	logger.warning(
	f"Expected to find noise in slice {slice}, but data has significant variation with std {acq_std}."
	)

[Copilot]

h5py is imported at module import time, but the Python package/conda recipe in this repo doesn’t list h5py as a dependency. This makes the converter unusable in a standard install (import fails immediately). Consider moving the h5py import inside convert() (or guarding it with a try/except that raises a clear install hint), and/or add h5py to the package/conda runtime dependencies.

[Copilot]

These assert statements are being used for validating user input / file structure (header content). Asserts can be disabled with python -O, and they also produce unhelpful error messages for CLI users. Prefer explicit validation with a raised exception (e.g., ValueError/RuntimeError) including what was expected vs. what was found.

[Copilot]

New acquisitions don’t set head.encoding_space_ref. Other generators/converters in this repo set it explicitly (e.g., python/mrd/tools/phantom.py:142-146, python/mrd/tools/ismrmrd_to_mrd.py:372-396) so downstream tooling can reliably map acquisitions to header.encoding[...]. Consider setting head.encoding_space_ref = 0 when synthesizing acquisitions here for consistency.

[Copilot]

np.transpose(...) returns a non-C-contiguous view; the MRD binary writer falls back to element-by-element serialization for non-contiguous arrays, which can be extremely slow for image-sized arrays. Consider making the transposed array contiguous (e.g., np.ascontiguousarray(...)) before writing.

Suggested change

	data = np.expand_dims(data, axis=(0, 1))
	data = np.expand_dims(data, axis=(0, 1))
	data = np.ascontiguousarray(data)

[Copilot]

Typo in comment: “tranpose” should be “transpose”.

Suggested change

	# So we need to tranpose X/Y, and add empty channel and slice dimensions
	# So we need to transpose X/Y, and add empty channel and slice dimensions

[Copilot]

validate_input_file() is currently a no-op and convert() re-implements validation inline. Either implement validate_input_file() and call it from convert(), or remove the stub to avoid dead code / confusion.

Suggested change

	def validate_input_file(f: h5py.File) -> None:
	"""Ensure the input fastMRI file contains the expected datasets."""
	def convert(input_filename: str, output_data_filename: str | None, output_images_filename: str | None) -> None:
	"""Convert fastMRI HDF5 file to MRD format."""
	with h5py.File(input_filename, "r") as f:
	# First validate the input file
	required_datasets = [FASTMRI_DATASET_NAME_HEADER]
	if output_data_filename is not None:
	required_datasets += [FASTMRI_DATASET_NAME_KSPACE]
	if output_images_filename is not None:
	required_datasets += [FASTMRI_DATASET_NAME_RECON_RSS]
	for dset_name in required_datasets:
	if dset_name not in f:
	raise RuntimeError(f"Input file is missing required dataset '{dset_name}'")
	def validate_input_file(f: h5py.File, require_kspace: bool, require_images: bool) -> None:
	"""Ensure the input fastMRI file contains the expected datasets."""
	required_datasets = [FASTMRI_DATASET_NAME_HEADER]
	if require_kspace:
	required_datasets.append(FASTMRI_DATASET_NAME_KSPACE)
	if require_images:
	required_datasets.append(FASTMRI_DATASET_NAME_RECON_RSS)
	for dset_name in required_datasets:
	if dset_name not in f:
	raise RuntimeError(f"Input file is missing required dataset '{dset_name}'")
	def convert(input_filename: str, output_data_filename: str | None, output_images_filename: str | None) -> None:
	"""Convert fastMRI HDF5 file to MRD format."""
	with h5py.File(input_filename, "r") as f:
	# First validate the input file
	validate_input_file(
	f,
	require_kspace=output_data_filename is not None,
	require_images=output_images_filename is not None,
	)