Artificial Intelligence Versus Human Interpretation in Diagnostic Radiology: A Systematic Review of Diagnostic Accuracy and Operational Efficiency.
Authors
Affiliations (3)
Affiliations (3)
- Medicine, Navarra University Foundation (UNINAVARRA), Neiva, COL.
- Medicine, Faculty of Health Sciences, Surcolombiana University, Neiva, COL.
- Research and Innovation Center (CIINA), Navarra University Foundation (UNINAVARRA), Neiva, COL.
Abstract
Rising imaging workload and a shortage of radiology specialists have driven interest in artificial intelligence (AI) to improve diagnostic speed and accuracy. This systematic review, without meta-analysis, describes how AI affects diagnostic accuracy and operational efficiency compared with human interpretation in diagnostic radiology. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) and Cochrane guidance, PubMed, ClinicalKey, ScienceDirect, and Google Scholar were searched on 11 April 2026 for studies published from January 2019 onwards. Two reviewers independently performed study selection, data extraction, and a Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2)-based risk-of-bias assessment. Twenty-one studies were included: 11 compared standalone AI with human readers, six compared AI-assisted with unassisted readers, and four reported both; 18 evaluated task-specific image-analysis AI and three evaluated generative AI models. Because no subgroup was homogeneous enough to pool, we synthesised results narratively and stratified them by comparison type, AI category, imaging modality, and study quality. Standalone image-analysis AI achieved high sensitivity for fractures and selected chest findings but performed less well on subtle findings and lower-quality images, and in one computed tomography-referenced comparison a radiologist outperformed AI. AI assistance improved reader performance more consistently, particularly among less experienced readers. Generative models did not reach specialist performance on complex tasks. Efficiency gains were concentrated in normal studies, and the only randomised trial found lower mammography screen-reading workload with similar early screening performance. No study assessed patient outcomes, and 18 of 21 studies were retrospective. This evidence supports AI only as a supervised, complementary tool; improvements in diagnostic accuracy should not be equated with clinical benefit.