Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach

PMID: 24892406 · DOI: 10.1038/ncomms5006 · Journal: Nature Communications (2014)

TL;DR

Aerts et al. extracted 440 quantitative radiomic features (intensity, shape, texture, wavelet) from pre-treatment CT scans of 1,019 lung and head-and-neck cancer patients across seven independent cohorts. They built a fixed four-feature radiomic signature on the Lung1 NSCLC training set (n=422), then validated it without retraining in three independent cohorts (Lung2 NSCLC, H&N1 and H&N2 HNSCC). The signature, dominated by intratumour-heterogeneity descriptors, was significantly prognostic in all validation sets (concordance index 0.65–0.69), outperformed or complemented TNM staging, and was associated with cell-cycling gene-expression programs in the Lung3 radiogenomics cohort (n=89). The work established radiomics as a noninvasive, low-cost, transferable prognostic phenotype across two cancer types. PMID:24892406

Cohort & data

• Total: 1,019 patients with NSCLC or HNSC head-and-neck cancer across seven CT image cohorts. PMID:24892406
• Training (NSCLC): Lung1 — 422 NSCLC patients treated at MAASTRO Clinic (Maastricht, NL); CT scans, manual delineations, clinical and survival data. Publicly available on TCIA as NSCLC-Radiomics. PMID:24892406
• Validation (NSCLC): Lung2 — 225 NSCLC patients treated at Radboud University Medical Center (Nijmegen, NL). PMID:24892406
• Validation (HNSCC): H&N1 — 136 HNSCC patients treated at MAASTRO Clinic. PMID:24892406
• Validation (HNSCC): H&N2 — 95 HNSCC patients treated at VU University Medical Center (Amsterdam, NL). PMID:24892406
• Radiogenomics (NSCLC): Lung3 — 89 NSCLC patients with pretreatment CT plus Affymetrix HuRSTA_2a520709 gene expression (21,766 genes). Publicly available on TCIA as NSCLC-Radiomics-Genomics. PMID:24892406
• Stability (test–retest): RIDER — 31 NSCLC patients with two CT scans ~15 min apart. PMID:24892406
• Stability (inter-observer): Multiple delineation — 21 NSCLC patients delineated by five oncologists each. PMID:24892406
• Assay / method: 440-feature CT radiomics feature library implemented in Matlab; survival modelling via Cox proportional hazards using the 4-feature radiomic signature. PMID:24892406

Key findings

• Univariate prognosis (Lung1-derived median thresholds, no retraining): 238 / 440 features (54%) significant in Lung2, 135 / 440 (31%) in H&N1, 186 / 440 (42%) in H&N2; 66 features (15%) significant in all three validation cohorts (G-rho test, FDR 10%). PMID:24892406
• Four-feature signature (built on Lung1, locked weights): (I) Statistics Energy (overall density), (II) Shape Compactness, (III) Grey Level Nonuniformity (intratumour heterogeneity), (IV) wavelet Grey Level Nonuniformity HLH (heterogeneity at mid-frequencies). PMID:24892406
• Multivariate validation (concordance index): Lung2 CI=0.65 (P=2.91 × 10⁻⁹, Wilcoxon), H&N1 CI=0.69 (P=7.99 × 10⁻⁷), H&N2 CI=0.69 (P=3.53 × 10⁻⁶). PMID:24892406
• vs. tumour volume: Volume alone performed well, but the radiomic signature was significantly better; combining signature + volume beat volume alone in all data sets. PMID:24892406
• vs. TNM staging: Signature outperformed TNM in Lung2 and H&N2, comparable in H&N1; signature + TNM significantly improved over TNM alone in all three validation cohorts. PMID:24892406
• HPV in HNSCC: No significant association of the signature with HPV status (P=0.17, Wilcoxon, combined H&N1+H&N2). Signature retained prognostic performance in HPV-negative subgroup (CI=0.66, n=130, 76% of patients), indicating complementarity to HPV screening. PMID:24892406
• Treatment-stratified analysis: Signature retained prognostic performance within both radiation-only and concurrent chemoradiation subgroups, in lung and head-and-neck cohorts. PMID:24892406
• Unsupervised clustering of Lung1: Three radiomic clusters significantly associated with primary T-stage (P < 1 × 10⁻²⁰), overall stage (P=3.4 × 10⁻³), and histology (P=0.019, squamous overrepresented in cluster II); no association with N-stage (P=0.46) or M-stage (P=0.73). PMID:24892406
• Radiogenomics (Lung3, GSEA on C5/MSigDB GO collection, FDR ≤20%): All four signature features showed significant gene-set enrichment. The two intratumour-heterogeneity features (III and IV) were strongly correlated with cell-cycling pathways, suggesting that more heterogeneous CT phenotypes reflect higher proliferation. PMID:24892406
• Feature-selection insight: Features with higher test–retest and inter-observer stability ranks tended to also have higher prognostic performance, supporting stability-based feature selection as a principled prefilter. PMID:24892406

Genes & alterations

This study used bulk gene expression (Affymetrix HuRSTA_2a520709, 21,766 genes) on the Lung3 cohort (n=89) only for gene-set enrichment analysis against radiomic feature ranks; no individual gene-level alterations or biomarker calls are reported. The principal biological readout is pathway-level: intratumour-heterogeneity radiomic features correlate with cell-cycling / proliferation gene sets. PMID:24892406

Clinical implications

• Prognostic biomarker: The four-feature radiomic signature is a noninvasive, low-cost prognostic marker in NSCLC and HNSCC patients treated with curative-intent radiotherapy or chemoradiation, computable from routine pretreatment CT without normalization. PMID:24892406
• Complementary to TNM: Adding the radiomic signature to TNM significantly improved prognostic discrimination in all validation cohorts, suggesting utility for risk stratification beyond anatomic staging — particularly relevant for non-resectable NSCLC patients receiving primary chemoradiation, where TNM is least informative. PMID:24892406
• Cross-disease transferability: A signature trained only on NSCLC generalized to HNSCC, supporting the existence of a “general prognostic tumour phenotype” detectable on CT and motivating broad applicability. PMID:24892406
• HPV-negative HNSCC: Signature retains prognostic value in HPV-negative head-and-neck patients, the higher-risk majority subgroup, suggesting added value to HPV-only stratification. PMID:24892406

Limitations & open questions

• Imaging variability: Authors acknowledge inter-hospital variability in CT acquisition; their analysis used raw DICOM data without correction or normalization. They argue this strengthens generalizability but anticipate further gains from standardized acquisition (e.g., NIH Quantitative Imaging Network, QIBA). PMID:24892406
• Image-quality differences: Lung CT (free-breathing, motion-affected) had relatively more noise than head-and-neck CT (head immobilization), which the authors propose as one reason the lung-derived signature actually performed better in head-and-neck cohorts. PMID:24892406
• Modality scope: Validated only on CT in two cancer types (NSCLC, HNSCC); transferability to MRI, PET, and other tumour types is hypothesized but not demonstrated here. PMID:24892406
• Single signature, single architecture: Only one fixed four-feature Cox model was tested in validation (a deliberate anti-overfitting choice); whether richer models or modern machine-learning architectures would substantially outperform was not assessed. PMID:24892406
• Patient selection: Cohorts restricted to curative-intent treatment with confirmed primary tumour; patients receiving palliative treatment or with prior lung/H&N cancers were excluded, limiting inference about advanced or pretreated disease. PMID:24892406
• Radiogenomics depth: GSEA was performed on a single gene-expression cohort (n=89) and only against the GO C5 collection at FDR ≤20%; mechanistic interpretation of heterogeneity ↔︎ proliferation is correlative. PMID:24892406

Citations from this paper used in the wiki

• “we present a radiomic analysis of 440 features … extracted from computed tomography data of 1,019 patients with lung or head-and-neck cancer.” (Abstract) PMID:24892406
• “The radiomic signature had good performance on the Lung2 data (CI=0.65, P=2.91 × 10⁻⁰⁹, Wilcoxon test), and a high performance in H&N1 (CI=0.69, P=7.99 × 10⁻⁰⁷, Wilcoxon test) and H&N2 (CI=0.69, P=3.53 × 10⁻⁰⁶, Wilcoxon test).” (Results — Prognostic validation of radiomic signature) PMID:24892406
• “The resulting radiomic signature consisted of (I) ‘Statistics Energy’ … (II) ‘Shape Compactness’ … (III) ‘Grey Level Nonuniformity’ … and (IV) wavelet ‘Grey Level Nonuniformity HLH’…” (Results — Building prognostic radiomic signature) PMID:24892406
• “both intratumour heterogeneity features in the signature (Feature III and IV) were strongly correlated with cell cycling pathways, indicating an increased proliferation for more heterogeneous tumours.” (Results — radiogenomics) PMID:24892406
• “We did not find a significant association between radiomic signature prediction and HPV status … (P=0.17, Wilcoxon test)… However, we found that the signature preserved its prognostic performance in the HPV-negative group (CI=0.66), consisting of the majority of patients (76%, n=130)…” (Results — HPV) PMID:24892406
• “The Lung1 data set, consisting of CT images for 422 NSCLC patients, and the Lung3 data set consisting of CT images and gene-expression profiling for 89 NSCLC patients, are publicly available at The Cancer Imaging Archive…” (Methods — Data sets) PMID:24892406

This page was processed by crosslinker on 2026-05-04.