Encrypted login | home

Program Information

BEST IN PHYSICS (JOINT IMAGING-THERAPY): Identification of Molecular Phenotypes by Integrating Radiomics and Genomics


P Grossmann

P Grossmann1*, O Grove2 , N El-Hachem3 , E Rios Velazquez1 , C Parmar1 , R Leijenaar4 , B Haibe-Kains5 , P Lambin6 , R Gillies2 , H Aerts1 , (1) Dana-Farber Cancer Institute/Harvard Medical School, Boston, MA, (2) H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, (3) Institut de Recherches Cliniques de Montreal, Montreal, Quebec, (4) Research Institute GROW, Maastricht, Netherlands, (5) University Health Network, Toronto, Ontario

Presentations

TU-CD-BRB-2 (Tuesday, July 14, 2015) 10:15 AM - 12:15 PM Room: Ballroom B


Purpose: To uncover the mechanistic connections between radiomic features, molecular pathways, and clinical outcomes, to develop radiomic based predictors of pathway activation states in individual patients, and to assess whether combining radiomic with clinical and genomic data improves prognostication.

Methods: We analyzed two independent lung cancer cohorts totaling 351 patients, for whom diagnostic computed tomography (CT) scans, gene-expression profiles, and clinical outcomes were available. The tumor phenotype was characterized based on 636 radiomic features describing tumor intensity, texture, shape and size. We performed an integrative analysis by developing and independently validating association modules of coherently expressed radiomic features and molecular pathways. These modules were statistically tested for significant associations to overall survival (OS), TNM stage, and pathologic histology.

Results: We identified thirteen radiomic-pathway association modules (p < 0.05), the most prominent of which were associated with the immune system, p53 pathway, and other pathways involved in cell cycle regulation. Eleven modules were significantly associated with clinical outcomes (p < 0.05). Strong predictive power for pathway activation states in individual patients was observed using radiomics; the strongest per module predictions ranged from an intra-tumor heterogeneity feature predicting RNA III polymerase transcription (AUC 0.62, p = 0.03), to a tumor intensity dispersion feature predicting pyruvate metabolism and citric acid TCA cycle (AUC 0.72, p < 10⁻⁶). Stepwise combinations of radiomic data with clinical outcomes and gene expression profiles resulted in consistent increases of prognostic power to predict OS (concordance index max = 0.73, p < 10⁻⁹).

Conclusion: This study demonstrates that radiomic approaches permit a non-invasive assessment of molecular and clinical characteristics of tumors, and therefore have the unprecedented potential to cost-effectively advance clinical decision-making using routinely acquired, standard-of-care imaging data. We show that prognostic value complementary to clinical and genomic information can be obtained by radiomic strategies.


Contact Email: