Click here to


Are you sure ?

Yes, do it No, cancel

Identifying Predictive Radiomic Markers for Patients in RTOG0617 Using Multiple Feature Selection Methods

N Sasankan1*, H Geng2 , H Zhong3 , Y Fan4 , M Rosen5 , Y Xiao6 , (1) ,Philadelphia, ,(2) University of Pennsylvania, Philadelphia, PA, (3) University of Pennsylvania, Philadelphia, PA, (4) University of Pennsylvania, Philadelphia, PA, (5) ,Philadelphia, ,(6) University of Pennsylvania, Philadelphia, PA


(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 6

Purpose: The multi-institutional RTOG 0617 study compares the treatment responses of high-dose (74Gy) vs standard dose (60 Gy) radiotherapy for non-small cell lung cancer. The patients enrolled in high-dose cohort had lower survival rate compared to the low-dose cohort. Though multiple factors may have contributed, we perform an analysis on radiomic data extracted for targets from pre-treatment DICOM images. A prognostic model is built based on these radiomics features that may help to identify patients who may benefit from a higher dose.

Methods: We analyzed 184 DICOM images of patients in the high-dose cohort. A total of 1130 feature were extracted using Pyradiomics.Multiple feature selection techniques were used to reduce the number of features and identify prognostic radiomic markers. We ultimately select MARS and MRMR method for feature selection.

Results: Two models using Boosted tree and Ensemble Bagged trees have produced models with reasonable accuracy using features from the MARS method. The Boosted method exhibited the following: accuracy, 67.4%; AUC, 0.68; specificity, 63%; sensitivity, 70%. The Ensemble Bagged trees methods performed with: accuracy, 69%; AUC, 0.72; specificity, 62%; sensitivity, 71%. Similarly Boosted trees and normal tree method performed well with the features identified from the MRMR method. The Boosted trees exhibited the following: accuracy, 70.7%; AUC, 0.74; specificity, 61%; sensitivity, 78% and the normal tree method exhibited the following characteristics: accuracy, 70.1%, AUC, 0.75; specificity, 66%; sensitivity, 73%. The accuracy of all models was evaluated using 50-fold cross-validation

Conclusion: Using the MARS method, the number of identified features were reduced to 9 from 1130. We noticed that the top 10 features produced by the MRMR method had the best overall accuracy. Predictive models built with these reduced features demonstrated reasonable performance.

Funding Support, Disclosures, and Conflict of Interest: This project was supported by NCI grants U24CA180803(IROC), U10CA180868(NRG), and PA CURE grants.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email