Evaluating Contrast-Dependent Errors in Virtual Contrast-Enhanced Brain MRI

Virtual contrast-enhanced MRI aims to synthesize post-contrast T1-weighted images from non-contrast scans. Evaluating synthesized T1CE is challenging because enhancement may occupy only a small part of the brain, while most anatomy remains similar before and after contrast. Global similarity metrics can therefore remain high even when enhancement is missed, misplaced, or accompanied by unwanted changes elsewhere. We address this by evaluating the contrast-induced change directly and separating change outside the expected enhancing region, recovery of enhancement magnitude, and spatial localization. We study these quantities using a conditional flow-matching model trained on BraTS Meningioma and evaluated on held-out BraTS-MEN data and BraTS-Africa. A threshold selected on BraTS-MEN validation data remains suitable on BraTS-MEN test data but identifies too little measurable enhancement on BraTS-Africa for broad enhancement-specific evaluation. Selecting a separate threshold using BraTS-Africa annotations restores measurement support, but enhancement recovery and localization remain weak. These findings show that image-level similarity alone does not establish whether the contrast-dependent signal has been recovered, and that the evaluation procedure itself must be examined when models are applied to substantially different clinical data.