Article Figures & Data
Figures
- FIG 1.
Overview of the methodology and overall workflow. MR imaging scans (Gd-T1w, T2w, and FLAIR) were preprocessed and segmented into 3 tumor subcompartments (necrosis, enhancing tumor, and peritumoral edema). Next, a set of radiomic features was obtained for each of the MR imaging scans across the 3 tumor subcompartments. The most stable and discriminatory features were identified using the PI score and AUC measures across every subcompartment and every MR imaging protocol. The identified stable and discriminatory features were used for stability assessment across tumor subcompartments (with the associated habitat), individual imaging protocols (Gd-T1w, T2w, FLAIR), and MR imaging acquisition parameters (magnetic field strength, pixel resolution, and slice spacing).
- FIG 2.
The strategy used for segregating our cohort into training (T̂ 1 –T̂ 4), validation (VS1–VS4), and TSs. The training and the validation sets were created using the LOSO scheme.
- FIG 3.
2D plots of discriminability (average AUC, y-axis) versus instability (PI score, x-axis) for each tumor subcompartment (A, Necrosis. B, Enhancing tumor. C, T2/FLAIR hyperintensity). Each radiomic feature is shown as a bubble where its size represents the SD between the AUC values across different LOSO runs. D, The stacked barplot of the total number of extracted features, the features that were identified as stable (PI = 0) from across different feature families per MR imaging sequence (Gd-T1w, T2w, and FLAIR), and the overlapping stable features across multiparametric MR imaging sequences. Different colors represent different feature families.
- FIG 4.
Barplots showing training and validation AUCs to distinguish patients with improved and poor outcomes using the top 5 stable and discriminatory radiomic versus discriminatory-only features. Blue barplots depict the AUC values using discriminatory features, while orange bars represent AUC values using stable and discriminatory radiomic features from the respective training and validation cohorts.
Tables
- Table 1:
Summary of patient demographics and the most commonly varying MR parameters across our cohort curated from 5 different sites (S1–S5)
Imaging Source Site S1, MD Anderson Cancer Center, Houston, Texas S2, Ivy Glioblastoma Atlas Project S3, Henry Ford Hospital, Detroit, Michigan S4, University of California, San Francisco, California S5, Cleveland Clinic, Ohio No. of cases 17 33 28 15 57 Magnetic field strength 1.5T 11 9 16 14 48 3T 6 24 12 1 9 Sex Female 7 17 5 6 23 Male 10 16 23 9 34 Scanner GE Healthcare 17 30 28 15 57 Siemens 0 3 0 0 0 Slice thickness (range) (mm) 6.5 0.9–6.5 1–6 1.5–6 0.9–7.5 Pixel spacing (range) (mm) 0.47–0.86 0.43–1.05 0.43–0.94 0.45–1.02 0.23–1 - Table 2:
A list of the top 5 features identified to reliably differentiate poor from improved survival in patients with GBM from the training set for every tumor subcompartment (necrosis, enhancing tumor, peritumoral T2/FLAIR hyperintensity, and nonenhancing tumor)
No. Feature Family Descriptor Window Size MRI Sequence Statistic Top 5 features selected from necrosis 1 Laws Edge-edge-level 5 T1 Skewness 2 Laws Level-level-level 3 T1 Skewness 3 Laws Wave-wave-level 5 FLAIR Median 4 CoLIAGe Sum variance 3 FLAIR Skewness 5 CoLIAGe Sum variance 3 FLAIR Kurtosis Top 5 features selected from enhancing tumor 1 CoLIAGe Sum average 3 T2 Median 2 CoLIAGe Sum average 5 T2 Median 3 CoLIAGe Energy 3 FLAIR Median 4 CoLIAGe Sum average 3 FLAIR Median 5 Laws Spot-level-ripple 5 T1 Skewness Top 5 features selected from edema and nonenhancing tumor 1 Laws Level-edge-level 5 T1 Median 2 CoLIAGe Sum variance 5 T2 Variance 3 CoLIAGe Sum variance 5 T2 Skewness 4 Haralick Sum entropy 5 FLAIR Skewness 5 Gabor –θ = 1.178, XZ –θ = 0, λ = 1.482, BW = 1 5 T2 Kurtosis Note:—BW indicates bandwidth.
Jump to section
Related Articles
Cited By...
- No citing articles found.