---

name: mri-operators description: "Use fastMRI complex tensor math, centered FFT/IFFT, RSS coil combine, SSIM loss, crops, normalization, and metric shape checks safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

fastMRI MRI Operators

Use this sub-skill when a task needs fastMRI tensor operators rather than data loading, model architecture design, or file-level leaderboard evaluation.

Operator Checklist

• Use fastMRI's real/imaginary tensor convention: complex PyTorch tensors are real tensors whose final dimension has size 2.
• Use fastmri.fft2c(data, norm="ortho") and fastmri.ifft2c(data, norm="ortho") for centered 2D transforms; spatial dimensions are -3 and -2, and complex real/imag is -1.
• Use fastmri.complex_abs, fastmri.complex_abs_sq, fastmri.complex_mul, fastmri.complex_conj, and fastmri.tensor_to_complex_np for complex tensor math.
• Use fastmri.rss(data, dim=...) for real-valued coil images and fastmri.rss_complex(data, dim=...) for complex coil data; always set the coil dimension explicitly when it is not dimension 0.
• Use fastmri.data.transforms.center_crop, complex_center_crop, center_crop_to_smallest, and normalize_instance for crop and normalization steps before losses or metrics.
• Use fastmri.SSIMLoss for tensor losses and fastmri.evaluate metrics for array-level reporting only after checking shape, crop, and finite-value assumptions.

References

• API reference explains exact operator signatures, tensor shape conventions, crop behavior, losses, metrics, and validation snippets.
• Troubleshooting maps common FFT, complex dimension, RSS, crop, SSIM, metric, dtype/device, and dependency failures to fixes.
• Smoke script runs a tiny CPU-only check for complex absolute value, centered FFT/IFFT shape, RSS, and center crop; use it when validating an environment or debugging operator imports.

Boundaries

• For reading k-space from fastMRI HDF5 files and applying masks in dataset transforms, use data-loading.
• For neural network architectures, training modules, VarNet, U-Net, and Lightning modules, use model-architectures.
• For file-based reconstruction evaluation, saving submissions, and challenge output directories, use evaluation-submission.