Corruption Structure Determines Failure Mode in MRI Reconstruction

Standard evaluation of learning-based MRI reconstruction focuses almost entirely on random undersampling, leaving open how reconstruction fails under the broader range of physically realistic measurement perturbations that arise in clinical practice. We present a controlled evaluation framework grounded in k-space physics, where the structure of any measurement corruption directly shapes reconstruction failure. Applying three physically motivated k-space corruptions to knee MRI from the fastMRI dataset (Zbontar et al., 2018), we show that corruption structure determines the qualitative character of failure independently of severity level. Random undersampling produces gradual degradation (normalized SSIM drop 0.134); structured line dropout produces catastrophic collapse with high cross-subject variance (normalized SSIM drop 0.877); and additive noise produces smooth, intermediate degradation (normalized SSIM drop 0.675). Notably, structured line dropout exhibits a high-variance regime at mild severity where visually similar corruption levels produce dramatically different failure outcomes across subjects, a non-obvious failure pattern absent from other corruption types. At severe severity, line dropout and additive noise both produce substantial degradation yet qualitatively distinct failure modes that aggregate PSNR evaluation cannot distinguish. These results argue for corruption-type-aware evaluation as a necessary component of reliable AI-based MRI reconstruction.