FARD — Frequency Attention Residual Denoising

Weights for FARD, the multi-contrast fusion network from Multi-Contrast MRI Acceleration via Post-Reconstruction Fusion (Medical Image Analysis, in press).

FARD refines vendor-reconstructed magnitude images from complementary orthogonal phase-encoding acquisitions. It is reference-free: it needs no fully-sampled high-resolution contrast to guide it, and no access to raw k-space. Each contrast is acquired with a reduced phase-encoding matrix along a different axis, so the three inputs lose resolution along different directions and jointly retain the spatial-frequency content needed to restore all of them.

Files

Three checkpoints, one per output contrast. Each takes all three accelerated contrasts as input and predicts its own target.

File	Output contrast	Params
fard_t1n.safetensors	T1-weighted	0.504 M
fard_t2w.safetensors	T2-weighted	0.504 M
fard_flair.safetensors	T2-FLAIR	0.504 M

These are weights only — no optimizer or scheduler state. checksums.json carries the SHA-256 of each file and the MD5 of the original .pth it was converted from, so provenance against the source repository is checkable.

Architecture

A 2D slice-wise network: 10 convolutional layers at C=64 with 3×3 kernels, block-wise gated attention at layers 2, 5 and 8, and frequency-domain processing at layers 3, 6 and 9. Residual connections every three layers. The three accelerated inputs are concatenated early and projected to C channels.

Where the parameters sit:

Component	Params
backbone (10 conv layers)	0.369 M
attention modules	0.084 M
frequency modules	0.041 M
fusion + aux conv + final layer	0.010 M
total	0.504 M

Usage

The architecture lives in the code repository; these files are the weights.

from safetensors.torch import load_file

model.load_state_dict(load_file("fard_t1n.safetensors"))  # strict=True

Channel order matters. Each checkpoint expects its own target contrast's accelerated input first:

Output	Input channel order
T1n	t1n, t2w, flair
T2w	t2w, flair, t1n
FLAIR	flair, t1n, t2w

Feeding a different order will produce plausible-looking but wrong output, since nothing in the network makes the ordering explicit.

Inputs

Bicubic-upsampled, vendor-reconstructed magnitude images, registered across contrasts and resampled to a common 1 mm³ grid. The code repository's data_prep/ reproduces that three-stage registration — intra-contrast, then inter-contrast to an anchor, then anchor to MNI.

Intended use and limitations

Research use only. This is not a medical device and has not been evaluated for clinical decision-making.

Trained and evaluated on brain MRI at approximately 4× protocol acceleration: a prospectively acquired 1.5T cohort, and retrospective simulation on BraTS-GLI 2024. Behaviour outside that regime — other anatomy, other field strengths, other acceleration factors, other vendors' reconstruction — is uncharacterised.

The network refines an existing reconstruction rather than reconstructing from k-space, so it cannot recover information the accelerated acquisition did not capture. On T1 in axial evaluation the degraded axis is through-plane, which limits what a slice-wise 2D model can restore.

Citation

@article{nazarov2026multicontrast,
  title   = {Multi-Contrast MRI Acceleration via Post-Reconstruction Fusion},
  author  = {Nazarov, Alexander and Kiryati, Nahum and Roizen, Dani and
             Kerpel, Ariel and Hoffmann, Chen and Greenberg, Gahl and
             Mayer, Arnaldo},
  journal = {Medical Image Analysis},
  year    = {2026},
  note    = {In press}
}

License

CC BY-NC 4.0 — non-commercial use, with attribution. This matches the license on the FARD architecture in the code repository; the surrounding pipeline there is MIT.