Program Information
Quantitative FLAIR MR Imaging as a Metric for MR Guided Radiation Treatment Planning (MRgRTP)
E Florez*, T Nicholas , S Lirette , A Fatemi , University of Mississippi Med. Center, Jackson, MS
Presentations
WE-F-205-8 (Wednesday, August 2, 2017) 1:45 PM - 3:45 PM Room: 205
Purpose: Significant variability in Glioblastoma multiform (GBM) radiation treatment margins around tumor and edema happen by difficulty to differentiate tumor vs edema. We used texture analysis technique on T2-FLAIR images to differentiate tumor vs. edema.
Methods: Eighteen GBM and ten meningioma or cerebral metastasis patients were selected. Neuroradiologist using semiautomatic algorithm to contour edema and tumor ROIs in all patients Figure 1. Texture Analysis (First-Order Statistics, Second-Order Statistics and Higher-Order Statistics) was applied to all ROIs, both with and without normalization. More than 280 different texture parameters were calculated for each ROI. The (least absolute shrinkage and selection operator) LASSO was applied on set of texture parameters to select one with highest association for distinguishing tumor vs edema. These variables used for ROC analysis and constructed all relevant plots.
Results: Variables selected for each scenario from the LASSO procedure are shown in Table 1. First-Order Statistics using 1%-Percentile feature was the only parameter chosen in all four scenarios with the best discriminant ability for meningioma, both with and without normalization. Second-Order Statistics using Correlation feature was also selected across scenarios, although the angle and the magnitude varied. For GBM, Higher-Order Statistics using Gray Level Non-uniformity and Short Run Emphasis features provided the best discrimination for images without and with normalization, respectively. Figure 2 displays ROC results showcasing both the single best discriminator and the discriminant ability of the model using all variables selected by LASSO. All univariate models had good discriminant ability (AUC>0.83), and all multivariate models had excellent discriminant ability (AUC>0.93). Figure 3 shows the sorted values for the best discriminator stratified by tissue.
Conclusion: Texture parameters, and only a small subset of such, show excellent ability to discriminate tumor vs edemas through its most discriminating features. In future we will include Diffusion-weighted images and ADC maps.
Contact Email: