2388
Gradient-entropy based radiomic features to predict molecular sub-types of pediatric Medulloblastoma on Gadolinium-enhanced T1w MRI1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 2Radiology, Children's Hospital Los Angels, Los Angels, CA, United States, 3Hematology, Children's Hospital Los Angels, Los Angels, CA, United States, 4Diagnostic Radiology, Cleveland Clinic Foundation, Cleveland, OH, United States
Synopsis
Genomic Characterization of Medulloblastoma (MB) has recently identified 4 distinct molecular subgroups: Sonic Hedgehog (SHH), wingless (WNT), Group 3, and Group 4. These subgroups have shown different clinical behaviours and benefits to subgroup-specific treatments. We explored the feasibility of a new gradient-entropy radiomic feature, CoLlAGe, to distinguish molecular sub-types of MB on Gd-T1w MRI. Our results using multi-class comparison via one way ANOVA and post-hoc comparison showed significant differences in CoLlAGe features obtained across molecular sub-types. Our feasibility results suggest that the CoLlAGe features in different tumor regions observed on routine Gd-T1w MRI may potentially serve as surrogate markers to non-invasively characterize molecular sub-types of pediatric MB.
Purpose
Medulloblastoma(MB) is the most common brain tumor in children1. Aggressive MB frequently requires cranio-spinal radiation, which also causes significant morbidity in long term survivors. Unfortunately, it is difficult to accurately determine how aggressive MB is in most patients. Hence, a ‘one-size-fits-all’ radiation and chemotherapy treatment regimen is widely adopted, causing children to undergo highly aggressive and in some cases, inadequate radiation therapy. Genomic characterization of MB has recently identified 4 distinct molecular subgroups: Sonic Hedgehog (SHH), wingless (WNT), Group 3, and Group 4.2 These subgroups have shown different clinical behaviours and benefits to subgroup-specific treatments. For instance, WNT group is known to have the best prognosis, while Group 3 carries the worst prognosis with 5-year survival rates of 50%.3 This molecular categorization is only feasible via an invasive biopsy, which carries severe risks associated with brain impairment. In this work, we explore the feasibility of a gradient-entropy based radiomic feature called CoLlAGe4 to distinguish molecular sub-types of MB. CoLlAGe aims to capture subtle differences in heterogeneity on a per-voxel basis. CoLlAGe features in previous studies4 have been found to be over expressed for more heterogeneous and aggressive pathologies, as compared to the less aggressive phenotypes in adult brain tumors. Hence, in this work, we hypothesize that CoLlAGe measurements when captured from the different tumor sub regions, will distinguish between the 4 molecular sub-types of MB, potentially capturing the underlying differences in tumor heterogeneity as reflected on a molecular scale.Methods
Our cohort consisted of 40 retrospectively collected studies (age: 3 months to 16.5 years) obtained from the Childrens Hospital of Los Angeles (CHLA). The cohort consisted of N=14 cases for SHH subtype, N=2 cases for WNT sub-type, N=7 cases for Group 3, and N=17 cases for Group 4 sub-type.Our workflow first involved registration of the Gd-T1w sequences with age-wise atlases, followed by harmonizing protocol-specific image intensities to template distributions to account for intensity non-standardness5. Expert delineation of enhancing, and necrotic regions was performed on T1w MRI and of edema region on T2w, and FLAIR. CoLlAGe features were then extracted on a per-pixel basis for every region (edema, tumor necrosis, enhancing tumor, non-enhancing tumor), for each of the MRI protocols. 4 statistics, mean, standard deviation, skewness, and kurtosis, were obtained from the extracted CoLlAGe features for every tumor region. We then performed a statistical experiment that aimed at finding significant differences among the 4 molecular subtypes. A one-way ANOVA test was performed followed by a “multiple comparison of means" Post Hoc test, to identify if the means of any groups have any significant statistical differences across the different subtypes of MB.Results and Discussion
Statistically significant features were found from the enhancing tumor, necrotic core, as well as the non-enhancing tumor regions. In the enhancing tumor region, 2 features- Variance of CoLlAGe differential average (p = 0.0301, Figure 1a) and Variance of CoLlAGe inertia (p= 0.0328) found statistically significant differences between Group 4 and SHH with a 95% confidence interval. We also found statistical differences in the necrotic core between Group 3 and Group 4, as well as group 3 and SHH. The two most significant features were the Skewness of the average sum of CoLlAGe(p = 0.00015, Figure 1b) and Variance of CoLlAGe (p = 0.0016) , both showing differences between Group 3 with SHH and Group 4.We also found that the absolute entropy of the
most aggressive sub-type - Group 3 was the highest in this region. Similarly, significant features were found in the non-enhancing tumor region, which showed differences between SHH and Group 3, as well as SHH and Group 4. The two most significant features included Skewness of CoLlAGe entropy (p= 0.0019, Figure 1c) and energy (p= 0.00128). Conclusion
Clinical Implications
Identifying non-invasive markers on MRI will have therapeutic implications in personalizing treatment management of pediatric MB. While promising, our findings were limited by the small sample size. Acknowledgements
No acknowledgement found.References
[1] Taran, S., Taran, R., Malipatil, N., and Haridas, K. Paediatric MB: An updated review. West Indian Med J. 2016 Mar 16;65(2):363-368. (2016).
[2] Kool, M., Korshunov, A., Remke, M., and et al. Molecular subgroups of MB: an international meta-analysis of transcriptome, genetic aberrations, and clinical data of WNT, SHH, Group 3, and Group 4 MBs.
[3] Lastowska, M., Trubicka, J.andNiemira, M., and et al. MB with transitional features between Group 3 and Group 4 is associated with good prognosis. J Neurooncol 138: 231 (2018).
[4] Prasanna, P., Tiwari, P., & Madabhushi, A. (2014). Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe): Distinguishing Tumor Confounders and Molecular Subtypes on MRI. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2014 (pp. 73-80). Springer International Publishing.
[5] Madabhushi, A., &Udupa, J. K. (2006). New methods of MR image intensity standardization via generalized scale. Medical Physics, 33(9), 3426-3434.
Figures
Figure 1: Box plots for the most significant feature in the different Tumor Sub-Regions
a) Box plots represent difference between group 4 and SHH in the Enhancing Tumor region
b) Box plots represent difference between group 3 with group 4 and SHH in the Nerotic Core
c) Box plots represent difference between SHH with group 3 and group 4 in the Non-Enhancing Tumor region