MRI Super-Resolution Erases Small Real Lesions Rather Than Hallucinating Fake Ones — And Erasure Worsens With Slice Thickness

arXiv 2608.06311·low signalXBlueskyLinkedInCopy link

A study submitted 2026-08-06 (2608.06311) tests whether super-resolution preprocessing corrupts white-matter hyperintensity segmentation, using 1-mm isotropic FLAIR scans from 29 ADNI subjects with expert manual segmentations, degraded to simulated 3-mm and 5-mm through-plane acquisitions and upsampled by multi-contrast implicit neural representation, the self-supervised ECLARE model, and cubic interpolation. Measuring detection sensitivity, erasure rate, and hallucination rate under MARS-WMH (the most small-lesion-sensitive of four segmenters tested), the dominant failure was erasure of small real lesions, not hallucination, and it grew with slice thickness — though every reconstruction still beat the raw thick slice. ECLARE recovered small-lesion signal best at both thicknesses while the INR was no better than cubic interpolation, a useful negative result for anyone assuming neural upsampling dominates classical methods.