An AI model described in a preprint synthesizes peak-enhanced post-contrast breast MRI from paired pre-contrast slices. In an exploratory ablation, the authors selected the configuration they judged to offer the best overall balance between tumor-focused measures and global image quality.
The report is an arXiv version 1 preprint dated 26 Aug 2026. Its hyperparameter comparison used 10 axial development instances sampled from available training data.
How the synthesis works
The system, MAMA-FLUX.2, uses conditional latent flow matching, with the pre-contrast image acting as spatial conditioning for the peak-enhanced target. Its training objective combines global latent flow matching with supervision focused on tumor regions and on stable foreground tissue.
The challenge data were derived from the MAMA-MIA breast DCE-MRI dataset. They consist of paired pre-contrast and peak-enhanced post-contrast images represented as 2D z-score-normalized slices, with tumor masks used for regional supervision and evaluation.
To adapt the pretrained generative system, the authors used LoRA fine-tuning. This updates adapter parameters in the transformer's attention projections while leaving the VAE, text encoder and pretrained transformer backbone frozen. Separate LoRA adapters were trained for axial and sagittal slices, and a lightweight binary logistic-regression classifier selected the corresponding adapter at inference.
The chosen compromise
The selected settings used a LoRA rank of 64 and scaling of 64, a maximum MHA value of 25, a tumor-loss weight of 0.25 and a stable-foreground loss weight of 0.1. The adapter had 66,846,720 trainable parameters—1.72% of the 3,875,544,576 parameters in the frozen base transformer—and occupied 255 MiB.
To judge the output, the study combined tumor-focused measures with whole-image similarity and error measures. Its reported scores included DSC of 0.8722 ± 0.0716 and HD95 of 16.8431 ± 38.9849, alongside LPIPS of 0.0745 ± 0.0249, MSE of 0.2592 ± 0.3837 and tumor SSIM of 0.8610 ± 0.0702. The ± figures are standard deviations; confidence intervals and hypothesis tests were not reported.
The pattern was a trade-off rather than a clean sweep. The authors reported that moderate regional weighting gave the most favorable development-set balance: higher stable-foreground weighting favored MSE and LPIPS, while stronger tumor weighting favored tumor SSIM without delivering the best overall balance. The 0.25 tumor and 0.1 stable-foreground weights were reported as offering the strongest segmentation-oriented result.
What remains uncertain
Qualitative examples broadly matched those numerical trends. Most shown examples preserved overall breast anatomy and enhancement in spatially aligned locations, while residual errors were most apparent in strongly or heterogeneously enhancing regions. The qualitative assessment was based on presented examples, not an independent clinical evaluation.
The result still rests on a narrow development exercise. The hyperparameter ablation covered only 10 development cases; the model was trained slice by slice without enforcing volumetric consistency; and the intensity window, regional loss weights and plane-routing classifier were chosen empirically. Those limits mean the selected configuration may not generalize beyond the studied setting.
The authors interpret the findings as indicating that conditional latent rectified-flow modeling is a promising formulation for pre-to-post contrast breast MRI synthesis, particularly when paired with tumor-aware and anatomy-preserving supervision. The supplied evidence remains an exploratory model-development result rather than a broader validation.
Funding and disclosures
The work received PLGrid computing facilities and support under grant PLG/2026/019392, along with partial support from the Excellence Initiative Research University program at AGH University of Krakow. The authors declared no competing interests relevant to the article.
Paper data and sources
Original title: MAMA-FLUX.2: Image-to-Image Synthesis of Post-Contrast Breast DCE-MRI for the MAMA-SYNTH Challenge
Authors: Kamil Kwarciak, Marek Wodzinski
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text