Artificial intelligence versus radiologists in predicting lung cancer treatment response: a systematic review and
Nehemias Guevara Rodriguez1, Noemy Coreas Mercado2, Kumar Panjiyar3
1Department of Medicine, Division of Hematology, Oncology and Bone Marrow Transplant, Saint Louis University School of Medicine, St. Louis, MO, United States.
Background:
Artificial intelligence (AI) has emerged as a promising adjunct to radiologist interpretation in oncology imaging. This systematic review and meta-analysis compares the diagnostic performance of AI systems versus radiologists in predicting lung cancer treatment response, focusing solely on treatment response rather than diagnosis.
Methods:
We systematically searched PubMed, Embase, Scopus, Web of Science, and the Cochrane Library from inception to March 31, 2025; Google Scholar and CINAHL were used for citation chasing/grey literature. The review protocol was prospectively registered in PROSPERO (CRD420251048243). Studies directly comparing AI-based imaging analysis with radiologist interpretation for predicting treatment response in lung cancer were included. Two reviewers extracted data independently (Cohen's κ = 0.87). We pooled sensitivity, specificity, accuracy, and risk differences using DerSimonian-Laird random-effects models. Heterogeneity (I²), threshold effects (Spearman correlation), and publication bias (funnel plots, Egger's test) were assessed. Subgroups were prespecified by imaging modality and therapy class.
Results:
Eleven retrospective studies (n = 6,615) were included. Pooled sensitivity for AI was 0.9 (95% CI: 0.8-0.9; I² = 58%), specificity 0.8 (95% CI: 0.8-0.9; I² = 52%), and accuracy 0.9 (95% CI: 0.8-0.9; pooled OR = 1.4, 95% CI: 1.2-1.7). Risk difference favored AI by 0.06 for sensitivity and 0.04 for specificity. AI's advantage was most apparent in CT and PET/CT, with smaller/non-significant gains in MRI. Egger's test suggested no significant publication bias (p = 0.21).
Conclusion:
AI demonstrates modest but statistically significant superiority over radiologists in predicting lung cancer treatment response, particularly in CT and PET/CT imaging. However, generalizability is limited by retrospective study dominance, incomplete demographic reporting, lack of regulatory clearance, and minimal cost-effectiveness evaluation. Prospective, multicenter trials incorporating explainable AI (e.g., SHAP, Grad-CAM), equity assessments, and formal economic analyses are needed.
Systematic Review Registration:
https://www.crd.york.ac.uk/prospero/, identifier CRD420251048243.
