The potential in artificial intelligence-driven radiomic signature to predict survival in patients with metastatic colorectal cancer treated with cetuximab-based therapy.

Abstract

247 Background: This analysis was undertaken to forecast survival and enhance treatment decisions for patients (pts) with colorectal cancer (CRC) with liver metastases sensitive to folinic acid, fluorouracil and irinotecan (FOLFIRI) alone [F] or in combination with cetuximab [FC] using simple quantitative radiomic changes between CT scans at baseline and 8 weeks. Methods: We retrospectively analyzed 667 pts with KRAS-unselected metastatic CRC in NCT00154102 treated with F and FC. CT quality was classified as high (HQ) or standard (SQ), and four data sets were created and named by treatment quality. Pts were randomly assigned 1:2 to training or validation sets: FCHQ, 38/78 pts; FCSQ, 62/124 pts; FHQ, 51/78 pts; FSQ, 78/158 pts. A machine-learning signature was trained using data set FCHQ to classify pts as treatment-sensitive or treatment-insensitive using just 4 of 3,499 potential radiomic imaging features. Performance was calibrated/validated using ROC curves. Hazard ratios (HRs) and Cox regression models were used to evaluate association with overall survival (OS). Results: The signature used decrease in tumor heterogeneity plus boundary infiltration to successfully predict sensitivity to FC (FCHQ: AUC, 0.80; FCSQ: AUC, 0.72) but failed with non-cetuximab regimens (FHQ: AUC, 0.59; FSQ: AUC, 0.55). The radiomic signature outperformed existing biomarkers ( KRAS mutational status and tumor shrinkage by RECIST 1.1) for sensitivity to cetuximab-based therapy and was strongly associated with OS in the cetuximab-containing sets FCHQ (HR, 44.3; p = 0.0001) and FCSQ (HR, 6.5; p = 0.005). Conclusions: This signature, derived from simple radiomic analysis of tumor imaging phenotype using only standard-of-care CT scans, appeared to be treatment-specific and was superior to all tested prognostic biomarkers. The signature provided early prediction of sensitivity and survival and could be used to guide treatment continuation decisions.