Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

Head and neck cancers

2042 - Development of a predictive radiomics signature for response to immune checkpoint inhibitors (ICIs) in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (RM-SCCHN)


08 Oct 2016


Head and neck cancers


Amy Prawira


Annals of Oncology (2016) 27 (6): 328-350. 10.1093/annonc/mdw376


A. Prawira1, P. Dufort2, J. Halankar2, D.M. Paravasthu2, A. Hansen3, A. Spreafico3, A.R. Abdul Razak3, E. Chen3, R.W. Jang3, U. Metser2, L.L. Siu3

Author affiliations

  • 1 Department Of Medical Oncology And Hematology, Princess Margaret Hospital, M5G 2M9 - Toronto/CA
  • 2 Joint Department Of Medical Imaging, University Health Network, Toronto/CA
  • 3 Department Of Medical Oncology And Hematology, Princess Margaret Hospital, Toronto/CA


Abstract 2042


Early phase clinical trials of ICIs in RM-SCCHN have shown promising results, but there is no validated predictive marker of response to date. We hypothesize that baseline host immune recognition creates a distinct microenvironment that is captured in computed tomography (CT)-images. Radiomics uses advanced image processing techniques to extract a large set of quantitative texture and geometric features from tumor regions of interest (ROI), and subject these to a machine learning protocol to train a classifier, which we exploit to develop as a predictive signature of response to ICIs.


We performed a retrospective analysis of clinical data and CT-images from prospectively enrolled cohorts of RM-SCCHN patients treated with ICIs in our institution. Tumor ROIs were manually contoured from baseline and first on-treatment CT-images. Extracted and computed radiomics features were employed to train a radial basis function support vector machine classifier to discriminate responders from non-responders. Ten-fold cross-validation protocol was employed to determine classifier accuracy.


Forty-two target lesions were contoured from 15 patients: median age = 58, males = 80%, p16-positive = 9, p16-negative = 5, p16-unknown = 1. Primary site: oropharynx = 11, oral cavity = 2, hypopharynx= 2. Treatment: anti-PD-L1 monotherapy = 8, combination of anti-PD-L1 and anti-CTLA-4 = 7. Excluded lesions: clinical data prematurity = 11. Per-lesion radiological outcomes: 20 responders, 11 non-responders. The radiomics classifier trained on the first on-treatment CT data achieved an accuracy of 79% (65.1% sensitivity, 88% specificity, p = 0.029) with an area under the ROC curve of 0.75. The classifier trained on the pre-treatment baseline CT data achieved an accuracy of 70.8% (37.9% sensitivity, 91.5% specificity, p = 0.16) with an area under the ROC curve of 0.60.


This pilot study showed early promising results. Patient accrual is ongoing, and further improvement in the accuracy of the developed algorithm is expected with increasing patient numbers.

Clinical trial identification

Legal entity responsible for the study

Amy Prawira


Princess Margaret Cancer Centre, Drug Development Program, Toronto, Canada.


L.L. Siu: Research funding from Merck to conduct clinical trials. All other authors have declared no conflicts of interest.

This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings