AI-Driven Radiomics in Precision Medicine

Abstract

Radiomics -- the high-throughput extraction and analysis of quantitative imaging features from medical images -- has emerged as a powerful tool for non-invasive tumour characterisation, treatment response prediction, and prognostication in precision oncology. This study developed and validated AI-driven radiomic models for three clinical prediction tasks: immunotherapy response in non-small cell lung cancer (NSCLC, n = 482), molecular subtyping of glioblastoma (GBM, n = 314), and recurrence prediction in hepatocellular carcinoma (HCC, n = 268) post-resection. A standardised radiomic pipeline extracted 1,218 features (shape, first-order, texture, wavelet) from CT (NSCLC, HCC) and MRI (GBM) using PyRadiomics v3.0, followed by LASSO feature selection and comparison of five classifiers (logistic regression, SVM, random forest, XGBoost, and a 3D-ResNet deep learning model). For immunotherapy response, the XGBoost radiomic model achieved AUC = 0.842 (external validation) -- significantly outperforming PD-L1 expression alone (AUC = 0.684; p < 0.001) and tumour mutational burden (TMB, AUC = 0.724; p = 0.008). For GBM molecular subtyping (IDH-mutant vs wild-type), the 3D-ResNet deep learning model achieved AUC = 0.924 on external validation. For HCC recurrence, the combined radiomic + clinical model achieved C-index = 0.782 -- significantly above clinical factors alone (0.648). These findings establish AI-driven radiomics as a clinically actionable precision medicine tool that extracts predictive information from standard-of-care imaging beyond what is accessible to human visual assessment.