MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Subregion-Based Radiomic Analysis of Preoperative Multi-Modal MR Images for Improving Glioblastoma Survival Outcome Prediction

J Fu1*, K Singhrao1, D Ruan1, X Qi1, J Lewis2, (1) Department of Radiation Oncology, UCLA, Los Angeles, CA, (2) Cedars-Sinai Medical Center, Beverly Hills, CA

Presentations

(Tuesday, 7/14/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: We hypothesized that subregion-based radiomic features could achieve more accurate survival outcome prediction in glioblastoma (GBM) patients than whole tumor-based features. However, manual GBM subregion segmentation is sensitive to large inter-observer variations and may result in unstable prediction models. In this study, we applied a two-stage automatic partitioning approach to identify the GBM subregions on multi-modal MR images and extracted the subregion-based features for predicting GBM patient survival outcome after gross total resection.


Methods: 101 GBM patients were included. Each patient had four MR images acquired before gross total resection. Tumor subregion contours were manually drawn by board-certified neuroradiologists. K-means clustering was first applied to generate patient-level supervoxels which were then fused by population-level hierarchical clustering to derive the autosegmented subregion contours. 1106 radiomic features were extracted from each subregion contour (manual and autosegmented) and whole tumor contour. The patient cohort was split into a good-outcome group of 50 patients and a bad-outcome group of 51 patients based on the median overall survival. Support vector machines (SVMs) based on three feature sets (whole tumor-based, manual subregion-based, and autosegmented subregion-based) were trained for outcome classification. Model performance was evaluated by computing the area under the receiver operating characteristic curve (AUC) using a 5-fold cross-validation protocol. Nested principal component analysis was used to reduce feature dimension to prevent model overfitting.


Results: SVMs trained using whole tumor-based features, manual subregion-based features, and autosegmented subregion-based features achieved the AUCs of 0.689 (95% CI: 0.586-0.792), 0.712 (95% CI: 0.608-0.817), and 0.723 (95% CI: 0.626-0.821), respectively.


Conclusion: Subregion-based radiomic features derived from manual or autosegmented contours of GBM subregions achieved more accurate outcome prediction than whole tumor-based features. Autosegmented subregion-based features achieved the best performance. Our study demonstrated the potential of using subregion-based features for improving GBM survival outcome prediction and assisting clinical decision-making.

Funding Support, Disclosures, and Conflict of Interest: This study was funded by Varian Medical Systems, Inc.

Keywords

MRI, Image Analysis, Feature Extraction

Taxonomy

IM- MRI : Radiomics

Contact Email