Radiomics Feature Extraction and Calcium Phenotype Discovery

A preliminary pipeline on 30 COCA patients achieves Spearman rho = 0.986 between extracted calcium volume and Agatston score, with 11/15 features significant after FDR correction. The project extends this into a full radiomics-based calcium phenotype discovery framework on the 789-patient Stanford COCA dataset. An endpoint-free validation framework is the core methodological contribution, necessary because COCA has no MACE outcomes. PyRadiomics (IBSI-compliant) extracts 80-120 raw features per patient, reduced to 50-80 robust features via ICC filtering and R2 hierarchical clustering. PCA eigen features are derived following Kolossvary et al., with consensus clustering across K-means, hierarchical, DBSCAN, and GMM. Validation uses Agatston stratification, perturbation ICC reproducibility, and clinical pattern alignment without outcome labels.