Answers are grounded in this paper and cite their sources.
Automated quality control in computed tomography begins by identifying which artifacts a reconstruction contains, because every correction is artifact-specific. Progress on this task is reported as a gain in F1 or exact-match accuracy, always measured under one choice of image normalisation. We show that this choice determines what the reported number means. We build a synthetic corpus in which beam-hardening, ring, metal-streak, and scatter operators are injected independently, so that the four-label presence vector is exact by construction and severity is recorded: 2,400 reconstructions split disjointly by phantom, with 360 more at severities outside the training range. A multi-label convolutional network reaches 0.9965 macro-F1 and 0.9860 exact-match accuracy, compared with 0.9811 and 0.9302, respectively, for gradient boosting on fourteen hand-designed features. Three findings transfer beyond this corpus. First, the feature-based models identify beam hardening and scatter from overall brightness rather than from cupping shape: two brightness features alone reach 0.797 macro-F1, and the network loses nothing when that cue is removed. Second, this shortcut conceals the confusion predicted by the underlying physics, since removing brightness raises beam-hardening/scatter confusion from 6.6% to 42.1%. Third, occlusion shows that the annotated metal region is neither necessary nor sufficient for detection, indicating that mask-overlap scores penalise a model for identifying the correct evidence.
No takes yet. Share an insight, caveat, or question.
Varshini et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: