Click here to


Are you sure ?

Yes, do it No, cancel

Investigating the Complementarity of Radiomics and Clinical Information for Predicting Treatment Failure in Multiple Cancer Types

M Vallieres1*, A Chatterjee2 , F Lucia3 , V Bourbonne3 , P Bonaffini4 , I Masson5 , A Mervoyer5 , C Reinhold4 , D Visvikis1 , U Schick3 , J Seuntjens2 , O Morin6 , M Hatt1 (1) Laboratoire de Traitement de l'Information Medicale (LaTIM, INSERM UMR 1101), Brest, France; (2) Medical Physics Unit, McGill University, Montreal, Quebec, Canada; (3) Radiation Oncology Department, CHRU Morvan Brest, Brest, France; (4) Radiology Department, McGill University Health Centre, Montreal, Quebec, Canada; (5) Radiation Oncology Department, Institut de Cancerologie de l'Ouest, Nantes, France; (6) Department of Radiation Oncology, Division of Physics, University of California San Francisco, San Francisco, CA, USA


(Thursday, 8/2/2018) 7:30 AM - 9:30 AM

Room: Davidson Ballroom B

Purpose: To investigate whether radiomics could complement baseline clinical information for improved prediction of treatment failure in cervical, meningioma and prostate cancers.

Methods: Cohorts of 191 cervical, 180 meningioma and 91 prostate cancer patients with PET and/or MRI pre-treatment scans were analyzed. For each cohort, data came from different institutions and/or from PET or MRI scanners with different characteristics. Wavelet filters were applied in 3D to all scans. For each filtered and image space, a total of 6095 radiomics features were computed by extracting textures with different parameterizations. Each cohort was randomly divided into 10 teaching and testing sets. The teaching sets were used to perform radiomics feature set reduction and optimize random forests models. The modeled endpoints were local failure (LF) and metastatic failure (MF) for cervical, LF for meningioma, and biochemical recurrence (BCR) for prostate cancer. Adjustments for class imbalance and institutional/scanner variability were applied during modeling. Internal training/validation splits in the teaching sets were used to estimate which teaching/testing split was most representative of a given cohort for reporting the testing performance via the area under the curve (AUC) and concordance-index (CI) metrics for binary and event-free-time endpoints, respectively.

Results: Models with ~10 variables were constructed for radiomics (i), clinical (ii) and combinations of both feature types (iii). In cervical cancer, the same performance was obtained for both LF and MF: (i) {AUC=0.82,CI=0.75}; (ii) {AUC=0.69,CI=0.66}; (iii) {AUC=0.83,CI=0.76}. In meningioma cancer, performance for LF reached: (i) {AUC=0.69,CI=0.71}; (ii) {AUC=0.76,CI=0.73}; (iii) {AUC=0.90,CI=0.91}. In prostate cancer, performance for BCR reached: (i) {AUC=0.76,CI=0.76}; (ii) {AUC=0.80,CI=0.78}; (iii) {AUC=0.69,CI=0.70}.

Conclusion: Our results suggest that LF and MF in cervical cancer are best modeled using radiomics variables only (non-significant improvement with clinical variables), LF in meningioma cancer using a combination of radiomics and clinical variables, and BCR in prostate cancer using clinical variables only.

Funding Support, Disclosures, and Conflict of Interest: Funding support: National Institute of Cancer (INCa project C14020NS), France


Texture Analysis, Modeling, CAD


IM/TH- Image Analysis (Single modality or Multi-modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email