Fig 1.
Diagram showing imaging process in radiomics.
(a) Multiple postprocessing steps were applied to magnetic resonance images, including image coregistration, adjustment of image resolution, and intensity normalization between patients. (b) The volume of interest covering the total tumor was manually drawn by a neuroradiologist. (c) Three categories of radiomics, namely histogram, three-dimensional geography, and textural analysis, were employed on the processed magnetic resonance images to yield radiomic features.
Fig 2.
Pediatric MBs were divided into four molecular subgroups.
(A) unsupervised NMF and (B) t-SNE analyses of the RNA-Seq gene-expression table. The 52 patients with MB were classified into WNT (7), SHH (17), Group 3 (14), and Group 4 (14). Abbreviations: NMF: Nonnegative matrix factorization; t-SNE: t-distributed stochastic neighbor embedding; WNT: Wingless; SHH: Sonic hedgehog; MB: Medulloblastoma.
Table 1.
Clinical profiles of the study cohort.
Fig 3.
Comparison of the radiomics features which were statistically different between 4 molecular subtypes of medulloblastoma.
In first 4 features (A-D) WNT and G3 groups tended to have the higher values than SHH and G4 groups. In the rest 4 features (E-H) SHH and G4 groups tended to have the higher values than WNT and G3 groups. (*: p<0.05; **: p<0.01).
Fig 4.
Comparison of survival in MB patients with different radiomics values.
For tumors with Energy value > 0.23 or SRLGLE value > 0.08, patients have shorter survival time than tumor don’t have these features (p = 0.04 in Log-rank [Mantel-Cox] test).
Fig 5.
Receiver operating characteristic analysis of the preliminary prediction model.
The prediction performance for predicting WNT, SHH, Group 3, and Group 4 MB was illustrated (area under the curve = 0.82, 0.50, 0.72, and 0.78, respectively).