Study: CT acquisition kernel and noise silently corrupt lung-nodule AI, undetected by DICOM metadata

A study by Daniel Soliman, published on arXiv (2606.12824), demonstrates that acquisition state—specifically CT reconstruction kernel and noise level—acts as a structured, measurable variable that silently corrupts lung-nodule AI performance, yet remains invisible to standard DICOM metadata. Using a LUNA16-trained MONAI RetinaNet detector, the researcher tested paired CT scans differing only in reconstruction kernel (B30f vs B80f) and found that kernel shifts alone altered AI-measured diameter and flipped a Fleischner size category in 5.2% (8/155) of nodules, while detection confidence was unchanged. Controlled perturbations on LIDC-IDRI data revealed a dissociation: the noise axis degraded detection confidence (p=5.9e-32) for nodules under 6mm, but did not affect measurement; conversely, the frequency/kernel axis corrupted measurement (p=8.6e-13) without impacting detection.

The study further shows that a simple 4-feature pixel fingerprint can recover reconstruction identity with high accuracy (patient-level AUC ~0.95 on real CT, 0.995 on a QIBA phantom), even where the ConvolutionKernel DICOM tag is identical across reconstructions. The kernel axis generalized across four manufacturers (leave-one-vendor-out AUC 0.94-0.98), matching within-vendor performance. This means the current ACR Assess-AI registry, which relies on DICOM metadata for context, misses critical input-side variations. The author argues that acquisition-aware, input-side validation is the missing layer for the ACR-SIIM Practice Parameter's acceptance testing and drift monitoring requirements now entering imaging-AI accreditation.