Abstract
Background: Radiomics enables the extraction of quantitative imaging biomarkers that can non-invasively capture tumor biology and treatment response. Delta-radiomics, by assessing temporal changes in radiomic features, may improve reproducibility and reveal early therapy-induced alterations. This study investigated whether delta-texture features from contrast-enhanced CT could predict progression-free survival (PFS) and overall survival (OS) in patients with metastatic colorectal cancer (mCRC) liver metastases treated with cetuximab rechallenge plus avelumab within the CAVE trial. Methods: This retrospective substudy included 42 patients enrolled in the multicenter CAVE phase II trial with evaluable liver metastases on baseline and first restaging CT. Liver lesions were manually segmented by two readers, and radiomic features were extracted according to IBSI guidelines. Delta-values were calculated as relative changes between baseline and post-treatment scans. Reproducibility (ICC > 0.70), univariate and multivariable analyses, ROC/AUC, bootstrap resampling, cross-validation, and decision curve analysis were performed to evaluate predictive performance and clinical utility. Results: Among reproducible features, delta-GLCM Homogeneity emerged as the most robust predictor. A decrease in homogeneity independently correlated with longer PFS (HR = 0.32, p = 0.003) and OS (HR = 0.41, p = 0.021). The combined clinical-radiomic model achieved good discrimination (AUC 0.94 training, 0.74 validation) and stable performance on internal validation (bootstrap C-index 0.77). Decision curve analysis indicated greater net clinical benefit compared with clinical variables alone. Conclusions: This exploratory study provides preliminary evidence that delta-GLCM Homogeneity may serve as a reproducible imaging biomarker of response and survival in mCRC patients receiving cetuximab plus avelumab rechallenge. If validated in larger, independent cohorts, delta-radiomics could enable early identification of non-responders and support personalized treatment adaptation in immuno-targeted therapy. Given the small sample size, the potential for overfitting should be considered. Future work should prioritize prospective multicenter validation with a pre-registered, locked model and explore multimodal integration (radiogenomics, circulating biomarkers, and AI-driven fusion of imaging with clinical/omic data) to strengthen translational impact. Beyond imaging advances, these findings align with broader trends in personalized oncology, including response-adaptive strategies, multimodal biomarker integration, and AI-enabled decision support.