Artificial intelligence for lung cancer: a systematic review of head‑to‑head CT, FDG PET/CT, and multimodal models across screening, staging, and prognosis.

backgroundArtificial intelligence (AI) has shown increasing potential in lung cancer imaging, particularly in detection, staging, prognosis, and recurrence prediction. However, there is limited synthesis of head-to-head comparative evidence between CT, FDG PET/CT, and multimodal fusion models within the same cohorts.

objectivesTo systematically review and critically appraise studies that directly compared CT-only, FDG PET/CT-only, and combined multimodal models in lung cancer, with emphasis on clinical setting, fusion strategy, validation design, and clinical utility.

methodsThis systematic review followed PRISMA 2020 and PRISMA-S guidelines. PubMed, Scopus, IEEE Xplore, and Google Scholar were searched for English-language human studies published between January 1, 2019, and September 8, 2025. Eligible studies reported same-cohort, head-to-head comparisons of CT, PET/CT, or multimodal models for lung cancer screening, staging, or prognosis. Risk of bias was assessed using PROBAST for prediction model studies and SANRA for narrative reviews. Data were extracted in duplicate and synthesized narratively, with meta-analysis performed where ≥ 3 studies were sufficiently homogeneous.

resultsFrom 2,417 records (PubMed 845, Scopus 920, IEEE Xplore 452, Google Scholar/manual 200), 31 studies met inclusion criteria (20 primary modeling studies, 11 reviews). In screening cohorts, low-dose CT deep-learning models consistently outperformed other modalities, with modest incremental value from clinical covariates. For nodal staging, integrated PET/CT radiomics-clinical models showed superior discrimination, calibration, and net-benefit compared with unimodal approaches. In prognostic and recurrence settings, fused PET/CT models outperformed CT- or PET-only models across institutions, with further improvement from clinical variables. Radiogenomics and pathology integration provided added value but were limited by small samples and lack of external validation.

conclusionsComparative evidence demonstrates that modality performance is context-dependent: CT dominates in screening, PET/CT fusion excels in staging and prognosis, and multimodal integration with clinical or biomarker data enhances discrimination and utility. Standardization, harmonization, and rigorous external validation remain critical for generalizability. CLINICAL TRIAL NUMBER: Not applicable.