Single and multi-site CT-based radiogenomics analysis of metastatic lung adenocarcinoma and correlations with outcome

Radiogenomic studies have mostly linked single-site radiomic features (RFs) to genomic alterations in locally-advanced lung cancer, limiting their applicability to patients with metastatic lung adenocarcinoma (MLUAD). Our aim was to evaluate associations between unsupervised CT-based radiomic clustering of single-site and multi-site features and oncogenic alterations (OAs) and response to treatment in MLUAD.

Patients managed at our center (October 2016-January 2024) with pre-treatment CT scans and next-generation sequencing were retrospectively included. Reproducible RFs were extracted from all solid tumor lesions > 1 cm³ using an automated pipeline. Patient-level integration used the centroid of each patient's lesions in radiomic space, providing multi-site radiomics data. RFs from the largest and biopsied lesions were also isolated. Patients were clustered by unsupervised hierarchical consensus clustering using centroid-based (Cluster-C), largest lesion (Cluster-M), and biopsied lesion (Cluster-B) features. Uni- and multivariable associations with OAs (any OA, smoker-related [sOA], non-smoker-related [nsOA], or wild-type), overall response rate (ORR), and overall survival (OS) were investigated.

Among 361 patients (median age 63.2 years; 41.3% women; 1721 segmented tumor lesions), 48.2% had sOA and 13% had nsOA. Cluster-M2 + M5 was enriched in KRAS (p = 0.048), MET (p = 0.046), and PI3KCA (p < 0.001) alterations. Cluster-M (especially Cluster-M2 + M5) independently predicted sOA (OR = 2.28, p = 0.006), and nsOA (OR = 5.49, p = 0.004). Cluster-M was linked to higher ORR (p = 0.026) and longer OS (p = 0.016).

Baseline CT-based single- and multi-site radiomics capture patterns associated with key OAs in MLUAD, suggesting their potential role as a non-invasive adjunct to guide molecular testing and optimize treatment selection.

Question In MLUAD, can single- and multi-site RFs from all measurable lesions enhance the detection of key OAs and outcome prediction beyond standard clinical-radiological assessment? Findings In 361 MLUAD patients, robust clustering using RFs from multiple tumor lesions per patient identified subgroups associated with key OAs, response to treatment, and survival. Clinical relevance Whatever the initial disease staging, radiomic clustering may serve as a non-invasive AI biomarker that complements molecular testing, helping identify actionable tumor profiles and stratify patients for treatment selection and prognostication in MLUAD.