pec-breast-mri-vce

PEC Breast MRI Virtual Contrast Enhancement

Original challenge artifacts for Predictive Enhancement Calibration for Latent Breast MRI Virtual Contrast Enhancement, accepted for an oral presentation at the MICCAI 2026 MAMA-SYNTH Challenge and recognized as a Top-3 challenge submission.

Paper · Code and reproduction guide · Interactive demo

News

• 2026-08-08: The interactive Hugging Face Space was released with paper-compatible float MRI inputs and public Yunnan examples.
• 2026-08-04: The verified original challenge weights were released here.
• 2026-08-01: Accepted for an oral presentation. Team tanlei086 achieved a Top-3 challenge submission.

Abstract

Virtual contrast enhancement (VCE) synthesizes enhanced breast MR images from pre-contrast acquisitions. Modern latent generators offer strong image priors, but their bounded natural-image autoencoders conflict with the non-canonical intensity scale of MRI. We show that the upper bound can alter radiomic fidelity before generation, while scaling source and target independently creates a coordinate inconsistency. We propose Predictive Enhancement Calibration (PEC), which represents each pair in a shared, case-adaptive coordinate during training and predicts its unavailable upper endpoint from the pre-contrast image at inference. We integrate PEC with a pretrained FLUX latent flow transformer via parameter-efficient reference conditioning. Target round trips first isolate representation loss before generation; near-matched conditional models then compare PEC with fixed-wide and separate coordinates under comparable training budgets and backbone settings. On the fixed internal MAMA100 development cohort, PEC improves all eight point estimates in this source-only VCE setting, with paired evidence strongest for MSE and LPIPS.

Challenge objective

<p align="center"> <a href="https://mamasynth.grand-challenge.org/"> <img src="assets/mama_synth_challenge_objective.png" width="100%" alt="MAMA-SYNTH challenge objective"> </a> </p> <p align="center"><em> Synthesize peak-enhancement breast MRI from a pre-contrast acquisition. Image courtesy of the MAMA-SYNTH organizers. </em></p>

Model description

PEC predicts a post-contrast intensity endpoint from a pre-contrast breast MRI slice and uses one shared, case-adaptive coordinate for conditioning and output decoding. The generator is a rank-128 reference-conditioning LoRA for black-forest-labs/FLUX.2-klein-base-9B. A fixed BF16 prompt embedding and the source-only PEC endpoint predictor are included for challenge-parity inference.

This repository is an adapter release. It does not contain the FLUX.2 transformer, VAE, text encoder, FP8 conversion, or medical images.

Files

lora/easycontrol_lora.safetensors   # final step-3156 rank-128 LoRA
pec/checkpoint_best.pt              # source-only PEC endpoint predictor
pec/config.yaml                     # portable predictor configuration
pec/normalization.json              # 47-token normalization statistics
prompt/prompt_da_peakpost_bf16.pt   # fixed challenge prompt embedding
encoding.json                       # shared PEC intensity contract
manifest.json                       # provenance and source/release hashes
assets/mama_synth_challenge_objective.png  # official task illustration
SHA256SUMS                          # release-file checksums

The release script verifies the archived source SHA-256 values before staging. Tensor values are unchanged; local filesystem paths in PyTorch/YAML metadata are replaced with portable identifiers.

Download

First accept the FLUX.2 [klein] Base 9B license, then download this adapter:

from huggingface_hub import snapshot_download

model_dir = snapshot_download(
    repo_id="QinLei086/pec-breast-mri-vce",
    local_dir="models/pec-breast-mri-vce",
)

Follow the inference guide to combine these artifacts with the official base model. Standard research inference uses BF16. The challenge's FP8 base transformer can be reconstructed locally from the licensed base weights; it is not redistributed.

Once data/processed/mama_source/mama100/source.jsonl has been reconstructed directly from public MAMA-MIA with scripts/build_mama_slice_manifest.py, the public recipe runs endpoint prediction, PEC source encoding, the released fixed prompt embedding, LoRA inference, and shared-coordinate decoding end to end:

bash recipes/infer_mama100.sh

Use PREPARE_ONLY=1 to verify endpoint prediction and source encoding without loading FLUX.2 9B. The recipe explicitly ignores target paths during endpoint prediction and strips target/mask fields from its inference manifest.

Training and evaluation

• Training data: tumour-positive DUKE and ISPY2 slices prepared through the MAMA-SYNTH pipeline; at most eight peak-lesion axial slices per patient.
• PEC predictor: 1,406 MAMA training patients plus 80 public Yunnan cases; 20 patient-disjoint Yunnan cases are held out for checkpoint selection. The 80/20 patient manifest and exact phase-verification table are public; the repository documents conversion from the raw Zenodo ZIPs.
• Development cohort: MAMA100.
• Generator: 20 epochs, rank 128, ROI-weighted flow loss.
• Inference: 30 steps, source-only PEC endpoint, shared intensity coordinate.

See the paper repository for the full data contract, configurations, metrics, and scripts.

Intended use and limitations

This release is for non-commercial medical-imaging research only. It is not a medical device and must not be used for diagnosis or treatment decisions. It was developed for two-dimensional, peak-phase breast MRI VCE and has not been clinically validated. Users must review TERMS.md and comply with the FLUX.2 license, dataset terms, and applicable acceptable-use policies.

Citation

@misc{lei2026predictiveenhancementcalibrationlatent,
  title         = {Predictive Enhancement Calibration for Latent Breast MRI
                   Virtual Contrast Enhancement},
  author        = {Qin Lei and Hao Wu},
  year          = {2026},
  eprint        = {2608.03612},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
  url           = {https://arxiv.org/abs/2608.03612}
}

Contact

Dr. Qin Lei<br> [email protected]