Program Information
Classification of Glioblastoma Multiforme Molecular Subtypes Using Three-Dimensional Multi-Modal MR Imaging Features
M Mulvey*, S Muhyadeen , U Sinha , San Diego State University, San Diego, CA
Presentations
SU-F-R-3 (Sunday, July 31, 2016) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Positive patient outcomes with the use of aggressive concurrent chemoradiotherapy have been shown for classical and mesenchymal glioblastoma multiformae (GBM) subtypes, while not having the same positive effect for patients with neural and proneural subtypes. Quantitative MR Imaging features from multimodal MRI have been recently explored for GBM subtype classification with some promising results. We have integrated 3D textural and moment invariant features to investigate if these will strengthen the association between glioblastoma MR image features and genomic markers.
Methods: 147 subjects with a complete set of T1, T1C (post-contrast T1), T2, and FLAIR imaging modalities without significant imaging artifacts acquired prior to surgery and with a known genomic subtype (Classical: 30, Mesenchymal:47, Neural:25, and Proneural:45) were chosen from TCIA database for GBM. The subject data were preprocessed and tumor and edema were contoured for the entire volume. Features were extracted from the segmented multi-modal data. These 3D features include volume, surface area, surface roughness, moment invariants, intensity statistics, and texture features. Feature selection and classification was performed using random forest. Features were extracted from the four modalities and difference fields of T1C-T1 and FLAIR-T2.
Results: Random forest out-of-bag classification error of 23.91% was achieved in the classification of mesenchymal group from the proneural group using the top twenty-eight selected features. These imaging features were primarily texture features on T1C but also include a moment invariant of T1, and ROI roughness statistics of the tumor and of the edema ROIs. These features show that tumor heterogeneity, gross shape, as well the tumor and edema border smoothness distinguish the molecular subtypes.
Conclusion: Using heterogeneous MRI data from the TCIA GBM database we have shown a 76% correct classification rate for GBM genomic subtypes with the same treatment protocol. Imaging features may lead to preoperative classification for improved treatment planning.
Contact Email: