Molecular profiling is essential in non-small cell lung cancer (NSCLC), but conventional genotyping depends on tissue or blood samples and can be limited by invasiveness, sample adequacy, turnaround time and accessibility. Radiomics and artificial intelligence could provide a complementary, non-invasive approach using routinely acquired CT images.

This systematic review and meta-analysis provides a comprehensive assessment of imaging-based models for predicting EGFR, ALK and KRAS driver mutations, including 124 studies and 45,747 patients, with 51 studies included in the quantitative analysis.

CT radiomics-based AI models showed promising performance for EGFR and ALK, both reaching a pooled sensitivity of 75.4%. Performance for KRAS was more heterogeneous and substantially lower, highlighting the need for further evidence.

For EGFR, adding clinical variables increased the overall AUC from 0.769 to 0.821, although the improvement was not conclusive and requires further validation.

Unlike previous reviews focused mainly on technical performance, this study takes a clinical perspective, comparing radiomics-only, clinical-only and combined models and focusing on CT, the standard imaging modality in NSCLC workflows. By including all eligible models rather than only the best-performing ones, it provides a broader and more realistic assessment of the available evidence.

The findings support AI-powered CT radiomics as a potential non-invasive screening tool to identify patients more likely to harbor targetable mutations and prioritize confirmatory molecular testing, rather than replace it.