Beyond noise: quality metrics as both confound and signal in clinical neuroimaging

In structural MRI, quality control (QC) metrics flag artifacts and reduce acquisition-related variance, but may also capture participant-related characteristics. We examined QC metrics from T1-weighted MRI in a transdiagnostic sample (N = 1,319; 865 controls, 454 patients across six groups). An independent split-half analysis identified a four-metric panel (SNR_total, CNR, iEFC, iIQR) spanning three dimensions (Signal, Contrast, Entropy/Intensity). Variance decomposition in healthy controls revealed distinct sources of variation: iIQR and CNR were predominantly acquisition-dependent, whereas iEFC was primarily participant-associated, particularly with sex, intracranial volume, and BMI. SNR_total showed a similar proportional split but explained less variance. Network analyses indicated that the metrics provide non-redundant information, while QC-demographic associations differed between controls and patients. These findings place QC metrics on a continuum from acquisition- to participant-associated variation: acquisition-dominant metrics may be appropriate technical covariates, whereas participant-associated metrics such as iEFC should not be routinely regressed out, particularly in case-control studies.