Program Information
Radiomics in CT Perfusion Maps of Head and Neck Cancer
M Nesteruk1*, O Riesterer1 , R Bundschuh2 , P Veit-Haibach1 , M Huellner1 , G Studer1 , S Stieb1 , S Glatz1 , M Pruschy1 , M Guckenberger1 , S Tanadini-Lang1 , (1) University Hospital Zurich and University of Zurich, Zurich, Switzerland, (2) University Hospital Bonn, Bonn, Germany
Presentations
SU-F-R-51 (Sunday, July 31, 2016) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose:
The aim of this study was to test the predictive value of radiomics features of CT perfusion (CTP) for tumor control, based on a preselection of radiomics features in a robustness study.
Methods:
11 patients with head and neck cancer (HNC) and 11 patients with lung cancer were included in the robustness study to preselect stable radiomics parameters. Data from 36 HNC patients treated with definitive radiochemotherapy (median follow-up 30 months) was used to build a predictive model based on these parameters. All patients underwent pre-treatment CTP. 315 texture parameters were computed for three perfusion maps: blood volume, blood flow and mean transit time. The variability of texture parameters was tested with respect to non-standardizable perfusion computation factors (noise level and artery contouring) using intraclass correlation coefficients (ICC). The parameter with the highest ICC in the correlated group of parameters (inter-parameter Spearman correlations) was tested for its predictive value. The final model to predict tumor control was built using multivariate Cox regression analysis with backward selection of the variables. For comparison, a predictive model based on tumor volume was created.
Results:
Ten parameters were found to be stable in both HNC and lung cancer regarding potentially non-standardizable factors after the correction for inter-parameter correlations. In the multivariate backward selection of the variables, blood flow entropy showed a highly significant impact on tumor control (p=0.03) with concordance index (CI) of 0.76. Blood flow entropy was significantly lower in the patient group with controlled tumors at 18 months (p<0.1). The new model showed a higher concordance index compared to the tumor volume model (CI=0.68).
Conclusion:
The preselection of variables in the robustness study allowed building a predictive radiomics-based model of tumor control in HNC despite a small patient cohort. This model was found to be superior to the volume-based model.
Funding Support, Disclosures, and Conflict of Interest: The project was supported by the KFSP Tumor Oxygenation of the University of Zurich, by a grant of the Center for Clinical Research, University and University Hospital Zurich and by a research grant from Merck (Schweiz) AG.
Contact Email: