MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Observer-Independent, Hand-Crafted Radiomic Features Predict GBM Patient-Specific Survival

E Carver1*, N Wen2 , E Liang3 , J Snyder4 , (1) Wayne State University, Troy, MI, (2) Henry Ford Hospital, Detroit, MI, (3) HFHS, Detroit, ,(4) HFHS, Detroit,

Presentations

(Sunday, 7/14/2019) 4:00 PM - 5:00 PM

Room: Exhibit Hall | Forum 2

Purpose: This study investigated radiomic features from pre-operative MR images to predict overall survival for patients with Glioblastoma(GBM). The focus was on tumor delineation as a factor in clinical utilization of radiomics by identifying overlapping statistically significant features from independently-created contours.

Methods: Pre-operative T1-weighted(T1W), T2-weighted(T2W), T1-contrast enhanced(T1CE), fluid attenuated inversion recovery(FLAIR) MR brain images, age at time of diagnosis, and overall survival were obtained from 73 GBM patients with varying levels of resection. Pre-processing included registration, resampling, and normalization. To investigate observer dependence, two independently created contour sets were created on these images. One contour was delineated by a physician and the other by GlistrBoost, a hybrid generative-discriminative model. Cancer Imaging Phenomics Toolkit(CaPTK) extracted 938 image features per sequence for each contour. Relevance of each individual feature was determined by showing significance in a univariate Cox Proportional Hazards(CPH) model, highly ranked by SVM recursive feature elimination(SVM-RFE) and finally possessing non-collinearity with other features. Only relevant features present in both contour radiomic feature datasets were included to establish interobserver independence. Collective predictive power of these overlapping relevant features was assessed using logistic regression(LR) and support vector machine(SVM) analysis with a validation dataset for both contours. Survival time was placed into three categories based on dataset patient survival time for regression analysis: <10, 10-15, and >15 months.

Results: 10 image features were found to be statistically significant on both sets of contours. LR/SVM showed the similarity of collective predictive power of these features on the validation dataset for feature data extracted from physician(0.64±0.15/0.56±0.18) and computer generated (0.63±0.16/0.56±0.16) contours.

Conclusion: While results are encouraging for establishing physician/machine contour independent image features, future research will take place to investigate age.

Funding Support, Disclosures, and Conflict of Interest: Research Scholar Grant, RSG-15-137-01-CCE from American Cancer Society

Keywords

MRI, Feature Extraction, Feature Selection

Taxonomy

IM/TH- Image Analysis (Single modality or Multi-modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email