Click here to


Are you sure ?

Yes, do it No, cancel

Development of Relapse-Free Survival Classifiers in Glioblastoma Patients by Integrating Multiparametric Magnetic Resonance Imaging Features

D Grass1 , K Chou1 , G Zhang1, C Liveringhouse2 , K Latifi1 , J Arrington1 , E Moros1*, M Yu1 , (1) Moffitt Cancer Center, Tampa, FL, (2) University of South Florida, Tampa, FL


(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Davidson Ballroom B

Purpose: Recurrence following standard therapy for glioblastoma (GBM) occurs in almost all patients. Improvement in recurrence prediction prior to therapy may enhance the ability of clinicians to recommend optimal treatment recommendation. This study was performed to build relapse-free survival (RFS) classifiers with machine learning approaches using habitat features extracted from diagnostic magnetic resonance images (MRI).

Methods: A total of 119 cases with first-time relapse following standard of care (maximal safe resection and chemoradiation) treatment were retrospectively studied. The cohort was randomly divided into training (n=80) and validation (n=39) groups. The median time to relapse in the training cohort was 276 days, which was used to dichotomize cases into long and short RFS. All cases had 5 MRI sequences (T1-post contrast, T1 weighted, T2 Flair, T2 weighted, ADC map) prior to treatment. An in-house program was applied to automatically extract habitat features from any combination of the 5 sets of images; a total of over 5000 features were generated. Otsu thresholding was applied to divide each image set into 2 intensity levels within gross tumor contours on images of T1-post contrast sequence. Machine learning software, weka, was utilized with logistic regression (LR), random forest (RF) and support vector machine (SVM) methods to generate classification. Attribute ranking was applied as the search method in both RF and SVM.

Results: LR did not generate a robust classifier. In the training group, RF provided 100% correct classification, and 67% in the validation. Using the SVM, the rates of correct classification were 95% and 67% respectively. The area under the receiver operating characteristic (ROC) curve was 0.62 and 0.68 in validation for the two classifications respectively.

Conclusion: A RFS prediction classifier for GBM built using 5-sequence MRI data is promising. Better prediction classifiers may facilitate design of personalized intervention.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email