Artificial Intelligence–Enabled Quality Control in Diagnostic Imaging: Evidence, Evaluation Requirements, and Human-Supervised Governance

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objective: To review AI-enabled quality control (QC) across diagnostic imaging modalities, distinguish algorithmic performance from transportability and clinical effectiveness, and propose an evaluation and governance framework for human-supervised QC agents. Methods: PubMed was searched in a targeted structured manner for English-language publications from January 2017 through August 4, 2026. Peer-reviewed studies were prioritized when they directly assessed acquisition quality, positioning and coverage, artifacts, protocol and dose, reconstruction, or equipment performance. Relevant professional statements were also traced. Findings were synthesized narratively, with SANRA items used for methodological self-checking; no meta-analysis or formal risk-of-bias assessment was performed. Results: Published applications map to five functions: examination assurance, acquisition assurance, output assurance, system assurance, and learning and governance. Evidence in radiography, CT, and MRI is dominated by retrospective, task-specific technical validation. Ultrasound and mammography include observational workflow studies after deployment. Patient-image evidence in nuclear medicine remains comparatively sparse, whereas phantom and equipment surveillance are closer to routine implementation. Recurrent limitations include inconsistent quality definitions, subjective labels, inadequate handling of leakage and clustering, limited external validation, and weak links between technical scores and diagnostic utility. Conclusion: AI can broaden QC coverage and accelerate feedback, but a universal technical score should not replace professional judgment. Human-supervised QC agents should coordinate specialized models and deterministic rules and should be constrained by human authorization, stage-gated validation, and life-cycle monitoring.