Reproduce “MAMA-FLUX.2: Image-to-Image Synthesis of Post-Contrast Breast DCE-MRI for the MAMA-SYNTH Challenge” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kamil Kwarciak, Marek Wodzinski
MAMA-FLUX.2 synthesizes peak-enhanced breast MRI images from pre-contrast scans by adapting a latent-flow generator with tumor-region and stable-foreground losses. In a 10-instance development ablation, reducing the stable-foreground weight from 1 to 0.1 raised mean Dice from 0.8263 to 0.8722 and lowered mean HD95 from 33.2027 to 16.8431, while MSE and LPIPS favored weight 1. The training-derived analysis does not establish clinical replacement.
Key insights
- Incorporating moderate regional supervision in the loss function (tumor-region and stable-foreground components) provides an optimal trade-off between global image fidelity and critical tumor-focused performance metrics.
- Increasing stable-foreground loss can improve general image similarity but may degrade segmentation-oriented metrics, highlighting the need for balanced supervision.
- Parameter-efficient fine-tuning via LoRA effectively adapts large generative models to specialized medical imaging tasks, requiring only 1.72% of the base transformer's parameters to be trained.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.