NIMG-38. QUANTITATIVE IMAGING PREDICTORS OF OVERALL SURVIVAL IN GLIOBLASTOMA PATIENTS ROBUST IN THE PRESENCE OF INTER-SCANNER VARIATIONS. (5th November 2018)

• Record Type: Journal Article
• Authors: Rathore, Saima; Bakas, Spyridon; Akbari, Hamed; Rozycki, Martin; Davatzikos, Christos
• Is Part Of: Neuro-oncology. Volume 20(2018)Supplement 6
• Page Start: vi184
• Page End: vi184
• Publication Date: 2018-11-05
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noy148.764
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: BACKGROUND: Glioblastoma is the most aggressive primary adult brain tumor with median overall survival (OS) of ~14 months following treatment. Although associations have been shown between multi-parametric magnetic resonance imaging (mpMRI) signatures and OS in glioblastoma patients, there is no sufficient validation across different institutions/scanners. This study explores appropriate normalization approaches and multivariate machine learning (ML) to identify robust and reproducible imaging predictors of OS across scanners/institutions. METHODS: We identified a retrospective cohort of 208 patients, who underwent surgery with gross total resection status, and had available mpMRI (T1, T1-Gd, T2, T2-FLAIR, DTI, DSC) data from three different scanners in the Hospital of the University of Pennsylvania. Median OS was used as a cut-off between long- and short-survivors. Inter-scanner harmonization of mpMRI was conducted by normalizing the tumor intensity profile, with that of the contralateral healthy tissue. Intensity distributions, morphological, statistical, and texture descriptors extracted from intensity-normalized tumor sub-regions (enhancing, non-enhancing, edematous), were multivariately integrated via ML to derive predictors of patient OS. The predictors generalizability on unseen patient data was evaluated under two configurations: i) pooled/scanner-agnostic using 10-fold cross-validation, and ii) across scanners (training in multiple scanners and testing in one). RESULTS: The accuracy for predicting long- versus short-survivors was 80.64% (sensitivity=82.82%, specificity=78.16%, area under the curve [AUC]=0.79) and 75.26% (sensitivity=80.80%, specificity=68.97%, AUC=0.75) for pooled/scanner-agnostic and across-scanner configuration, respectively. The short-survivors, compared to long-survivors, showed relatively large and irregular infiltrating tumor, increased and compromised tumor microvascularity (DSC, T1-Gd), lower water concentration (T2), and higher cell density (DTI). CONCLUSION: Our findings suggest that quantitative analysis of appropriately normalized clinically-acquired mpMRI, coupled with ML, yields non-invasive robust predictors of OS for glioblastoma patients in presence of inter-scanner variations. Further validation of our predictors in more extensive multi-institutional datasets, can facilitate their potential incorporation into clinical practice, influencing surgical decision-making, treatment planning, and assisting patient management.