You can deploy this model and start earning money today!

NKI-AI/
mirage-mama-synth

Name: mirage-mama-synth
Availability: InStock

Views

No views yet

MIRAGE — MAMA-SYNTH paper checkpoint

Official AIFOFOMO checkpoint for MIRAGE (Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement), the winning method in the MAMA-SYNTH challenge comparison.

Paper: Borghesi, A., Wang, X., Teuwen, J., and Yiasemis, G. MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement. arXiv:2607.19137, 2026.

Challenge: MAMA-SYNTH Grand Challenge — synthesise peak post-contrast 2D breast DCE-MRI from the matching pre-contrast slice (MAMA-MIA cohort).

MIRAGE architecture (Figure 1, arXiv:2607.19137)

Figure 1 — MIRAGE architecture. Residual U-Net with training-only auxiliary segmentation and frozen downstream nnU-Net guidance (discarded at inference).

MIRAGE qualitative synthesis examples (arXiv:2607.19137)

Qualitative examples — pre-contrast input, synthetic post (MIRAGE), and ground-truth post with ROI overlay (from the paper).

Files

File	Description
`model.pt`	Training checkpoint at iteration 59 000 (EXP-183; 80/20 train/val split)
`mirage.yaml`	AIFOFOMO training config from aifofomo (`projects/mama_synth/paper/mirage.yaml`)

Model

Architecture: residual 2D U-Net (mirage), 1→1 channels, features [64, 128, 256, 512]
Training-only aux heads: tumour segmentation (finest + deep supervision) and frozen nnU-Net downstream-seg loss — not used at inference
Inference I/O: pre-contrast slice in → synthetic post-contrast slice out (main_model_output)
Training losses (reference): masked L1/LPIPS, ROI under-enhancement, aux seg, downstream nnU-Net seg

Usage

This checkpoint is an AIFOFOMO / Lightning Fabric save. Load with the bundled mirage.yaml model block:

yaml

model:
  name: mirage
  in_channels: 1
  out_channels: 1
  features: [64, 128, 256, 512]
  residual: true
  deep_seg_supervision: true   # training only; inference uses image head only

For full training / validation reproduction, see the MAMA-SYNTH MIRAGE tutorial and projects/mama_synth/README.md.

Official challenge evaluation (FRD, AUROC, …) uses the mama-research/mama-synth evaluator.

Citation

bibtex

@article{borghesi2026mirage,
  title={MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement},
  author={Borghesi, Andrea and Wang, Xin and Teuwen, Jonas and Yiasemis, George},
  journal={arXiv preprint arXiv:2607.19137},
  year={2026}
}

NKI-AI/mirage-mama-synth