Kyrre Eeg Emblem MSc, PhD, Presenter: Intellectual property; NordicNeuroLab AS

Marco Cunha Pinho MD, Abstract Co-Author: Nothing to Disclose

Frank G. Zoellner, Abstract Co-Author: Nothing to Disclose

Paulina Due-Tonnessen MD, Abstract Co-Author: Nothing to Disclose

John K. Hald MD, Abstract Co-Author: Nothing to Disclose

Lothar R. Schad PhD, Abstract Co-Author: Nothing to Disclose

Torstein Meling, Abstract Co-Author: Nothing to Disclose

Otto Rapalino MD, Abstract Co-Author: Nothing to Disclose

Atle Bjornerud MSC, Abstract Co-Author: Intellectual property; NordicNeuroLab AS Board Member; NordicNeuroLab AS

To develop a generic support vector machine (SVM) model using MRI-based blood volume distribution data for preoperative glioma survival associations and to prospectively evaluate the diagnostic efficacy of this model in autonomous patient data.

Our study was approved by institutional and regional medical ethics committees. We retrospectively included 235 preoperative adult patients from two institutions with a subsequent histologically confirmed diagnosis of glioma after surgery. A SVM learning technique was applied to whole-tumor relative cerebral blood volume (rCBV) histograms from dynamic contrast enhanced MRI (1,2). SVM models with the highest diagnostic accuracy for 6-months, 1-, 2-, and 3-year survival associations were trained on 101 patients from the first institution. Using linear and cox regression analysis for diagnostic accuracy and survival associations, respectively, the diagnostic efficacy of the SVM models were tested on independent data from 134 patients from the second institution.

Compared to histopathology and presence of contrast enhancement, the whole-tumor rCBV-based SVM model was the strongest parameter associated with 6-months, 1-, 2-, and 3-year survival in the independent patient data (Chi-square = 25.49-48.43, P < 0.001; ROCAUC = 0.794-0.851). Results were corrected for known survival predictors, including patient age, tumor size, neurologic performance and postsurgical treatment.

Computer aided diagnosis in glioma survival analysis can reduce operator measurement errors (3). Our data show that SVM machine learning in combination with whole-tumor rCBV histogram analysis identifies early patient survival in gliomas regardless of traditional clinical and histopathological features. The SVM models presented are insensitive to patient- and institutional variations. (1) Boxerman JL, AJNR 2006; 27(4):859-67 (2) Emblem KE, JMRI 2013; doi:10.1002/jmri.24390. (3) Zacharaki EI, AJNR 2012; 33(6):1065-71  

Machine learning techniques have the potential to improve standardization of current advanced MRI methods for preoperative glioma characterization and from this aid treatment planning.

Emblem, K, Pinho, M, Zoellner, F, Due-Tonnessen, P, Hald, J, Schad, L, Meling, T, Rapalino, O, Bjornerud, A, Generic SVM Model for Preoperative Glioma Survival Associations: A Multi-center Validation Study.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14014831.html