Reducing False Positives and False Negatives in Artificial Intelligence-assisted Imaging Screening: A Critical Review of Accuracy, Error Mechanisms and Clinical Translation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Imaging-based screening programmes are constrained by two opposing failures: false-positive results that generate unnecessary recall, anxiety, further imaging and biopsy, and false-negative results that permit disease to progress undetected until the next screening round or until symptoms appear. Artificial intelligence applied to medical images has been presented as a means of relaxing this constraint by improving detection without expanding the recalled population. This critical review evaluates whether that claim is supported, and under what conditions. Evidence was drawn from peer-reviewed studies identified through open scholarly indexes and citation searching, with emphasis on prospective, randomised and population-based evaluations across breast, lung, ophthalmic, prostate and tuberculosis screening. Three findings emerge. First, the strongest current evidence, from randomised and large prospective breast screening studies, indicates that artificial intelligence used as triage plus decision support within a double-reading programme increases cancer detection and sensitivity while leaving specificity unchanged and substantially reducing reading workload; the effect on interval cancer rate is at present consistent with non-inferiority rather than with demonstrated superiority. Second, gains are configuration-dependent rather than intrinsic to the algorithm: autonomous rule-out of low-risk examinations has produced both preserved and significantly increased recall in different prospective settings, and simulated reader replacement has reduced detection in at least one population-based cohort. Third, residual error is structured rather than random, arising from shortcut learning, distribution shift, model version change, uneven subgroup performance and reader-algorithm interaction effects such as automation bias. The evidence base remains geographically narrow, dominated by breast imaging, largely reliant on retrospective and simulated designs outside mammography, and weakly connected to patient-relevant endpoints such as mortality, overdiagnosis and stage shift. Priorities include randomised evaluation outside mammography, prespecified subgroup and threshold reporting, standardised post-deployment monitoring of version change, and explicit accounting of the harms attached to each error type.