MIRAGE model boosts breast MRI contrast with lesion-aware AI

A new AI method called MIRAGE (Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement) addresses a challenging problem in breast imaging: inferring contrast-enhanced MRI from a single pre-contrast slice. Developed by Andrea Borghesi, Xin Wang, Jonas Teuwen, and George Yiasemis, MIRAGE uses a residual 2D U-Net architecture trained with three forms of lesion-aware supervision: an asymmetric penalty that heavily penalizes missed enhancements, multi-scale auxiliary tumor segmentation, and guidance from a frozen post-contrast tumor segmentation nnU-Net. This allows the model to focus on clinically relevant regions rather than optimizing pixel-level fidelity alone.

Tested on 301 cases from the multi-centre MAMA-SYNTH dataset, MIRAGE ranked first on six out of eight complementary metrics including image quality, regional accuracy, radiomics, and segmentation-based measures. It markedly improved downstream lesion localization compared to tuned pix2pix, conditional diffusion, and latent bridge-matching baselines. However, generative alternatives retained advantages in LPIPS (perceptual similarity) and contrast classification, revealing a clear fidelity-utility trade-off. The study demonstrates that task-aware synthesis is optimal only when defined by specific downstream models and metrics, offering important guidance for clinical AI deployment.