PEC method outperforms in breast MRI virtual contrast, besting FLUX baseline

Virtual contrast enhancement (VCE) generates contrast-enhanced breast MR images from pre-contrast scans, reducing the need for contrast agents. However, modern latent generators like FLUX use autoencoders trained on natural images, which clash with MRI's non-canonical intensity scale. This causes two problems: the bounded encoding can distort radiomic features before generation, and scaling source and target independently creates a coordinate mismatch.

Qin Lei and Hao Wu introduce Predictive Enhancement Calibration (PEC), which represents each pre-contrast/contrast-enhanced pair in a shared, case-adaptive coordinate during training. At inference, PEC predicts the missing upper endpoint from the pre-contrast image alone, aligning coordinates without requiring a target. They integrated PEC with a pretrained FLUX latent flow transformer using parameter-efficient reference conditioning. In experiments on the internal MAMA100 development cohort, PEC improved all eight evaluation metrics in source-only VCE, with the most significant gains in MSE and LPIPS. The work ranked top-3 in the MICCAI 2026 MAMA-Synth Challenge, and code is available online.