AI Integration in Lung Cancer Imaging: A Systematic Evidence Map and Scoping Review of Workflow Efficiency Outcomes, Time-to-Diagnosis, and Implementation Gaps

Background: Lung cancer remains the leading cause of cancer mortality globally, yet artificial intelligence (AI) tools deployed to accelerate pulmonary nodule detection have not demonstrated consistent translation of radiologist-facing efficiency gains into patient-facing reductions in time-to-diagnosis (TTD). We conducted a systematic evidence map and scoping review to quantify what workflow efficiency metrics have been studied, by whom, and what critical gaps persist. Methods: We systematically searched six open-access databases and preprint servers (MEDLINE via PubMed, Europe PMC, medRxiv, ClinicalTrials.gov, WHO ICTRP, and Cochrane CENTRAL) for English-language studies published between January 1, 2020, and April 16, 2026, evaluating AI integration in lung cancer imaging workflows (chest radiography [CXR] or computed tomography [CT]) that reported ≥1 workflow efficiency metric across four priority tiers: P1 (core workflow: TTD, reading time, workload reduction), P2 (system-level: turnaround time [TAT], critical finding notification), P3 (human factors: cognitive load, patient wait time), and P4 (downstream outcomes: emergency department [ED] length of stay, after-hours work). Risk of bias was assessed using ROBINS-I for non-randomized studies and RoB 2 for randomized controlled trials (RCTs). A visual evidence gap map was constructed to chart metric reporting density across the included studies. Results: Of 6,631 records identified, 13 studies met inclusion criteria. No single study reported more than 3 of 12 possible workflow metrics (median = 1). TTD—the most clinically meaningful endpoint—was reported in only 3/13 studies (23%), with no standardized definition across studies. Reading time was the most studied metric (6/13, 46%) but predominantly in simulated reader studies. Complete evidence deserts existed for time-to-critical-finding notification, worklist clearance rate, patient wait time, cognitive load/burnout, user interactions, and after-hours work reduction (0/13 each). The largest RCT (Woznitza et al., 2026; n = 93,326) demonstrated a 27.4% reduction in median CXR reporting time (34.1 h vs. 47.0 h, p < 0.001) but zero reduction in TTD (44 days vs. 46 days, p = 0.84)—directly challenging the assumed causal pathway. Overall ROBINS-I risk of bias was Moderate for 11/12 applicable studies, Serious for 1 (Hill et al., 2025), and Low for 1 (Lancaster et al., 2025). Conclusions: The AI-lung cancer imaging literature suffers from a profound outcome metric mismatch: extensive investment in radiologist-facing efficiency metrics with negligible measurement of patient-facing outcomes. The absence of mediation analyses, standardized TTD definitions, and human-factors outcomes represents a critical evidence gap that must be addressed before AI's clinical value in lung cancer pathways can be established. Keywords: artificial intelligence; lung cancer; time-to-diagnosis; workflow efficiency; scoping review; evidence map; radiology; computer-aided detection; implementation science; PRISMA-ScR Keywords: artificial intelligence; lung cancer; time-to-diagnosis; workflow efficiency; scoping review; evidence map; radiology; computer-aided detection; implementation science; PRISMA-ScR