Disentangling Lung-Cancer CT/LDCT AI: A Systematic Evidence Map of Clinical Tasks, Evidence Chains, and Translational Gaps

Detection dominates over prediction: Of 293 studies reviewed, 230 (78.5%) focused on detecting existing lung cancer rather than predicting future risk, meaning most AI research addresses identifying tumors that are already present rather than identifying which patients will develop cancer—a clinically distinct and arguably more valuable task.

Critical validation gaps limit clinical readiness: Only 29% of studies included external validation (testing on data from different sources), 20.5% reported calibration metrics, and 13% used decision-curve analysis. These are standard requirements for determining whether an AI tool will actually work in real clinical settings, suggesting most studies lack evidence of real-world applicability.

Explainability remains underdeveloped: Only 21.5% of studies included saliency or attribution methods (techniques that show why the AI made a decision), which is essential for clinicians to trust and understand AI recommendations, particularly in high-stakes medical decisions like cancer diagnosis.

Few studies complete the full evidence chain: Only 10.6% of studies (31 out of 293) completed all six tiers of evidence needed to translate research into clinical practice, with the greatest dropout occurring at the "reasoning/explanation" stage—indicating a substantial gap between published research and deployment-ready AI systems.

Longitudinal evidence is sparse: Only 17.7% of studies used longitudinal CT scans (follow-up imaging over time), which is necessary to validate whether AI predictions about future cancer risk actually materialize, limiting confidence in prognostic claims.