ESTRO 2020

Session Item

November 30
10:30 - 11:30
Physics Stream 2
Proffered papers 31: Quantitative imaging and radiomics
Proffered Papers
11:10 - 11:20
Prospective Validation of a Radiomics Signature for Chemoradiotherapy Lung Cancer Patients
Sean Walsh,


Prospective Validation of a Radiomics Signature for Chemoradiotherapy Lung Cancer Patients
Authors: Philippe Lambin.(University Maastricht, The DLab, Maastricht, The Netherlands), Ralph Leijenaar.(OncoRadiomics, Research and Development, Liege, Belgium), Benjamin Miraglio.(OncoRadiomics, Research and Development, Liege, Belgium), Akshayaa Vaidyanathan.(OncoRadiomics, Research and Development, Liege, Belgium), Sean Walsh.(OncoRadiomics, Research and Development, Liege, Belgium), Fadila Zerka.(OncoRadiomics, Research and Development, Liege, Belgium), Samir Barakat.(OncoRadiomics, Research and Development, Liege, Belgium)
Show Affiliations
Purpose or Objective

Radiomics refers to the comprehensive quantification of tumour phenotypes and is a promising field of scientific study with a large amount of activity in recent times. However, to date no clinical level of evidence 1 has been provided for any of the many radiomics signatures published in the literature. The purpose of this study is to provide that evidence by prospectively validating the original prognostic radiomics signature for chemoradiotherapy lung cancer patients (Aerts et al.).

Material and Methods

A cohort of 300 chemoradiotherapy stage I-III NSCLC patients were collected from the observational clinical study SDC lung (NCT01855191). The GTV within the CT scans of these patients was used as input to the radiomics signature (a combination of intensity, shape, texture and wavelet features). Delineations were performed manually as part of the routine clinical workflow. The primary outcome measure was two-year survival rate after treatment. The signature was used to classify patients as responders or non-responders in this context. Pre-specified statistical tests were performed to assess the performance of the signature.


KM curves visualize a clear split between groups classified as responders/non-responders (high/low prognostic score based on a median prediction threshold of the original signature of Aerts et al.) A log-rank test indicates a significant split (p-value = 0.002). Discrimination by the signature was assessed by calculating Harrell’s concordance metric (c-index = 0.64, 95%CI: 0.57-0.71). Cox regression was performed on the signature to determine the calibration slope (0.93) and a likelihood ratio test indicates a valid relative risk model as the slope is close to 1 and not significantly different from 1 (p-value = 0.820). Additionally, the coefficients of the individual variables of the signature were join tested in the validation cohort (p-value = 0.001) indicating that the performance in the prospective validation cohort could be improved by adjusting the original coefficients of the features.


To the best of our knowledge this study demonstrates the first clinical evidence level 1 (i.e. prospective study) for any radiomics signature. This has implications for the wider field as it demonstrates that other signatures could also be prospectively validated. This signature could be used as a clinical decision support tool to evaluate the likelihood of response of NSCLC patients treated by chemoradiotherapy. This could be practically used in treatment planning to justify additional emphasis on adding chemotherapy and/or optimizing the dose to the GTV in patients classified as a non-responder. Conversely, this could be used in treatment planning to justify additional emphasis on optimizing reduced dose to the OARs in patients classified as a responder.  Other potential applications of this signature include use as a stratification tool in future trials or evaluation of elderly patient for which we have no randomized trial outcome data.