Multimodal AI for early lung cancer detection: from LDCT to clinical and molecular integration

Low‑dose computed tomography (LDCT) is the established screening backbone for lung cancer in high‑risk populations. However, its clinical value depends on a complete pathway, including eligibility assessment, image acquisition, nodule interpretation, follow‑up, referral, quality assurance, and harm reduction, and not on image detection alone. Artificial intelligence (AI), radiomics, clinical‑risk models, and blood‑ or breath‑based biomarkers are being investigated as tools to improve decisions within that pathway. This narrative, state‑of‑the‑art translational review synthesizes evidence on AI‑enabled integration of clinical risk, LDCT‑derived imaging features, radiomics, longitudinal imaging, and molecular or experimental biomarkers for early lung cancer detection and pulmonary nodule triage. AI‑enabled integration may improve defined decisions within the LDCT pathway, most plausibly pulmonary nodule triage. To distinguish genuine decision support from simple data fusion, this review applies a five‑tier framework: data aggregation, diagnostic enrichment, risk stratification, threshold‑based triage, and pathway‑level decision support across all modalities and studies. Within this framework, cell‑free DNA (cfDNA) methylation and cfDNA fragmentomics occupy a higher translational tier than autoantibodies, generic serum markers, breathomics, or metabolomics, because their signals have a coherent biological basis and they have been tested in multimodal nodule‑triage studies with external validation. AI is best understood as an integrative decision layer within LDCT‑centered pathways. Its value depends on whether it supports calibrated, fair, and actionable decisions, not on discrimination alone. The guiding question of this review is not about a general comparison between AI and biomarkers. Rather, it is about identifying which data layer most benefits each particular decision, pinpointing the exact stage of the LDCT pathway where that decision occurs, and evaluating that benefit based on rigorous evidence of calibration, safety, and net clinical benefit.