Four-feature radiomic signature (Aerts 2014)
Overview
A locked four-feature Cox proportional hazards prognostic model built from CT radiomics features, derived by Aerts et al. (2014) on the Lung1 NSCLC training cohort (n=422) and validated without retraining in three independent cohorts. The four features are: (I) Statistics Energy (overall CT density), (II) Shape Compactness, (III) Grey Level Nonuniformity (intratumour heterogeneity), and (IV) wavelet Grey Level Nonuniformity HLH (mid-frequency heterogeneity). Validated concordance indices ranged from 0.65 (Lung2) to 0.69 (H&N1, H&N2). The signature is the canonical example of cross-cancer-type transferability of a radiomic prognostic marker.
Used by
- PMID:24892406 — four-feature radiomic signature trained on Lung1 NSCLC (n=422), validated in Lung2 NSCLC (CI=0.65, P=2.91×10⁻⁹), H&N1 HNSCC (CI=0.69, P=7.99×10⁻⁷), and H&N2 HNSCC (CI=0.69, P=3.53×10⁻⁶); outperformed or significantly complemented TNM staging in all validation cohorts; intratumour-heterogeneity features (III and IV) correlated with cell-cycling gene-expression pathways in the Lung3 radiogenomics cohort (n=89) PMID:24892406.
Notes
- The signature was intentionally fixed (locked weights) to avoid overfitting; only a single four-feature Cox model was evaluated in validation — a deliberate anti-overfitting design choice PMID:24892406.
- HPV status in HNSCC did not associate with signature score (P=0.17, Wilcoxon); the signature retained prognostic value specifically in the HPV-negative majority subgroup (CI=0.66, n=130) PMID:24892406.
- The mechanistic link between imaging heterogeneity and proliferation (cell-cycling pathways) is correlative; causal interpretation is not established PMID:24892406.
Sources
This page was processed by crosslinker on 2026-05-04.