---

name: evaluation-submission description: "Prepare fastMRI reconstructions for validation, metric evaluation, zero-filled baselines, v2 filename conversion, and leaderboard-style submissions." disable-model-invocation: true metadata: disco-role: operating license: MIT

fastMRI evaluation-submission

Use this sub-skill when an agent needs to create, validate, rename, evaluate, or explain fastMRI reconstruction outputs for validation/test-style submission workflows.

Do not use this sub-skill for HDF5 dataset loading internals, low-level MRI tensor operators, or training loops; route those to data-loading, mri-operators, or the training sub-skill instead.

References and scripts

• Evaluation reference: Use this for exact public APIs, metric definitions, challenge target keys, and evaluation crop behavior.
• Submission workflows: Use this for end-to-end output preparation, zero-filled baselines, metric runs, v2 filename conversion, and BART caveats.
• Troubleshooting: Use this to diagnose missing datasets, filename/key mismatches, empty filters, crop issues, unavailable leaderboards, and optional BART setup failures.
• Zero-filled reconstruction script: Use this bundled baseline script to turn fastMRI k-space HDF5 files into reconstruction prediction files without depending on repository example paths.
• Validation script: Use this before metrics or packaging to check prediction filenames, HDF5 keys, target compatibility, optional filters, and v2 suffix expectations.

Critical facts

• fastmri.save_reconstructions(reconstructions, out_dir) writes one .h5 per dictionary key and stores the array under dataset key reconstruction.
• fastmri.convert_fnames_to_v2(path) renames .h5 files missing the _v2.h5 suffix and raises ValueError if the directory path does not exist.
• Metric evaluation expects target files and prediction files to share filenames; prediction files must contain dataset reconstruction.
• --challenge multicoil uses target key reconstruction_rss; --challenge singlecoil uses target key reconstruction_esc.
• The evaluation path center-crops both target and reconstruction to the square target width before computing MSE, NMSE, PSNR, and SSIM.
• The public fastMRI leaderboard site is unavailable in this checkout's README notice, so prepare and validate outputs but do not promise upload availability.

Minimal commands

python scripts/validate_reconstructions.py --predictions-path PREDICTIONS --target-path TARGETS --challenge multicoil
python -m fastmri.evaluate --target-path TARGETS --predictions-path PREDICTIONS --challenge multicoil
python scripts/zero_filled_reconstruction.py --data-path KSPACE_DATA --output-path PREDICTIONS --challenge multicoil

For public knee v2-style filenames after writing predictions:

from pathlib import Path
from fastmri import convert_fnames_to_v2
convert_fnames_to_v2(Path("PREDICTIONS"))