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

CT image standardization is superior to larger but heterogeneous data for robust radiomic models

Date

05 Apr 2019

Presenters

Diem Vuong

Citation

Annals of Oncology (2019) 30 (suppl_2): ii20-ii21. 10.1093/annonc/mdz066

Authors

D. Vuong1, M. Bogowicz1, J. Unkelbach1, R. Foerster1, S. Denzler1, A. Xyrafas2, M. Pless2, S. Thierstein2, S. Peters3, M. Guckenberger1, S. Tanadini-Lang1

Author affiliations

  • 1 Radiation Oncology, University Hospital Zurich, 8091 - Zürich/CH
  • 2 SAKK - Swiss Group for Clinical Cancer Research, 3008 - Bern/CH
  • 3 Multidisciplinary Oncology Center, Centre Hospitalier Universitaire Vaudois - CHUV, 1011 - Lausanne/CH
More

Resources

Background

Radiomics is a promising tool for identification of new prognostic biomarkers. Radiomic models are often based on single-institution data however multi-centric data, highly heterogeneous due to different scanning protocols, reflect better the clinical reality. Robustness studies are crucial to find features robust to e.g. scanner settings. We study if a CT radiomics overall survival (OS) model trained on multi-centric data with robust feature pre-selection can achieve a similar performance as a model on standardized data.

Methods

Pre-treatment CT data from 121 IIIA/N2 NSCLC patients from a prospective multi-centric randomized trial (SAKK 16/00, neoadj. chemo- or radiochemotherapy prior to surgery) were used to calculate 1404 radiomic features on the primary tumor. Two OS radiomic models were trained on (1) a sub-cohort with standardized imaging protocol (native CT, standard kernel, n = 84) and on (2) the entire heterogeneous cohort but with robust radiomic feature pre-selection. Robust features were extracted from four robustness studies (contrast, convolution kernel, motion, delineation) using the intra-class correlation coefficient (> 0.9 considered stable). Seperately for each model, principal component (PC) analysis was performed and PCs describing in total 95% data variance were selected. The feature with highest correlation to the PCs were used for the multivariate Cox model with backward selection. Model performances were quantified using Concordance Index (CI), verified with 10-fold cross-validation and compared using bootstrap with resampling.

Results

Robustness studies revealed 113 stable features. The convolution kernel was the largest influence on the feature stability. Final OS model on the entire heterogeneous data consisted of four and on standardized data of six features (all identified as unstable). The model on standardized data showed significant better prognostic performance compared to the model with robust feature pre-selection based on the entire heterogeneous data (CI = 0.64 and 0.61, p < 0.05, resp.).

Conclusions

For our prognostic NSCLC radiomic models image protocol standardization appears superior to using larger but heterogeneous imaging data combined with robust feature pre-selection.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Swiss National Science Foundation (SNSF) Swiss Group for Clinical Cancer Research SAKK.

Disclosure

M. Guckenberger: Board member: European Society for Therapeutic Radiation Oncology (ESTRO). 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