Multiomic Approach to Radioresistance of Ependymomas in Children and Adolescents

The resistance to treatment prompted us to perform the national PEPPI study (Pediatric Ependymoma Photons Protons and Imaging). PEPPI collated clinical, imaging and dosimetry data from children with intracranial ependymoma treated in France between 2000 and 2013. Since then, patients are treated in the current prospective SIOP II Ependymoma program. Our clinical results confirmed the classical clinical prognostic factors including radiotherapy dose, and we showed that imaging biomarkers from T2/FLAIR, perfusion and diffusion MRI were novel prognostic factors. We reported that relapse after RT occurs mainly locally within the high-dose region. The prognostic value of dose and the high rate of relapse with standard dose prompted us to perform an in silico dosimetry study of dose escalation comparing photons and protons, confirming the feasibility of this approach. In PEPPI series, the prediction of the site of relapse with advanced MR imaging was not possible due to small sample of imaging data and the recent biomolecular classification was not available. Recently, DNA methylation profiling provided a novel classification of ependymoma in molecular subgroups harbouring a strong prognostic value.

EPENDYMOMICS project aims to determine the prognostic role of multimodal imaging, to identify radioresistant clusters predictive of relapse and to analyze the link between radiomics features and genomic markers, hence leading to a radiogenomics study. It will build on a database with a larger number of patients called NETSPARE (Network to Structure and Share Pediatric data to Accelerate Research on Ependymoma) to develop a radiomics approach, i.e. predict clinical endpoints by relying on quantitative features extracted from medical images using either handcrafted or automated features extraction algorithms. These features will be exploited and combined with other available variables (clinical, genetic, etc.) as improved decision support.

1. Data collection MRI: Diagnostic, FU until relapse (DICOM format)

• Sequences: T1W, T2W, FLAIR, T1W with \& without contrast enhancement, diffusion-weighted imaging (DWI), perfusion-weighted imaging (PWI) Radiotherapy: CT scan, RTDOSE, RTSS, RTPLAN (DICOM RT format) Clinical: Age, surgery, chemotherapy, doses, late effects, relapse date Histology: Hematoxylin and Eosin (H\&E)-stained histopathology slides Molecular Biology: Methylation groups RELA, YAP, PFA, and PFB

1. Data quality check and delineation of volume of interest to curate data before post-processing and analysis. To ensure preliminary robust image segmentation the Radiotherapy Structure Set will be checked for each patient. The referring radiologist will confirm all imaging changes after RT. An expert radiation oncologist will segment volumes of post-treatment abnormalities in T1WI post contrast, in T2W/FLAIR imaging, as well as site of relapse and missing OARs.
2. Imaging data post-processing and analysis of changes after radiotherapy We will process the DWI and PWI with the Olea Sphere®3.0 software, a post-processing solution for MRI and CT scanners. We will extract and generate Apparent Diffusion Coefficient maps from DWI data calculated on a voxel-by-voxel basis and relative cerebral blood volume maps calculated from PWI data with an oscillation-index singular value decomposition routine and correction for T1-weighted leakage effects. All the MRI data will be strictly co-registered with the T1WI-PC and planning CT.
3. Radiomics harmonization We will perform the increase of the level of harmonization of images and/or extracted features on highly heterogeneous dataset due to its multiple sources. We will develop a GAN-based framework, translating heterogeneous images to match the properties of a standard dataset, such as a template reference image, or alternatively, one set of images chosen as a reference. We need to determine the relevant properties within images (local or global metrics, texture, edges, contrast, signal-to-noise ratio, etc.) and to ensure the ability of the framework to harmonize images without losing their clinically relevant informative and content. In the features space, numerous statistical approaches can be applied, such as normalization or batch effect compensation. ComBat has been shown to outperform other similar approaches and we will rely on Monte Carlo estimation for small samples. We will investigate the combination of the batch-correction methods with unsupervised clustering to deal with data presenting very high heterogeneity and a very small number of samples per batch. It might be beneficial and complementary to combine image-based and feature-based harmonization methodologies for improving the results of multicenter radiomics studies. We will evaluate the potential added benefit of both previously developed approaches in improving the results. The goal of this task will be to establish whether the first or the second approach (or the combination of both) is the most efficient, taking into account not only the absolute improvement observed in the results, but also the computing time and effort required to implement each approach in practice.
4. Biomolecular analysis Biological resources from PEPPI and SIOPII ependymoma cohorts of patients will be included. First, the diagnosis of intracranial ependymomas will be confirmed and then pathological features will be investigated. WHO 2020 classification of ependymomas: the cIMPACT-NOW (Consortium to Inform Molecular and Practical Approaches to CNS Tumor Taxonomy - Not Official WHO) decided to classify ependymomas according to location and the underlying genetic alteration. The current WHO classification is under revision. Biological studies required to classify intracranial ependymoma of the project: Immunohistochemistry: p65-RelA, H3K27Me3. FISH analysis (classification and Copy Number Variations). DNA-methylation profiling. RNAseq analysis.
5. Deep learning (DL) for molecular biomarkers of ependymoma For this purpose, the Zimmer team has designed and trained deep convolutional neural networks based on the Inception V3 architecture to distinguish ependymoma from glioma grades III and IV. Preliminary results based on an imaging data set indicate a high accuracy of discrimination and suggest the possibility to predict clinically relevant mutations from H\&E images alone (manuscript in preparation). We will extend these DL approaches to predict resistance to therapy and metastasis. This will be done by using the histopathology images alone either as input or in combination with classifications based on molecular assays. At a technical level, we will use state-of-the-art deep neural network architectures based on convolutional layers and skip connections. To compensate for the moderate size of the training data set, we will make heavy use of transfer learning. Here, we will make specific use of the TGCA database, which contains images from very different cancers such as breast cancer, and adapt successful models by retraining on a subset of the ependymoma data acquired in this study. We also will adopt multi-scale image representations for a better exploitation of the image information at cellular and tissue levels, use H\&E specific color normalization and methods for automated elimination of regions subject to imaging artefacts. We also will incorporate the developed methods in Imjoy. By decoupling the graphical user interface from the computational back end, it enables users to access high-end computing capabilities from any workstation or laptop.
6. Statistical analysis Several signatures will be developed and validated : a radiomic signature (RAD-Score), a radiomic and clinical signature (RADClin-Score), a molecular signature (MOL-score) and a global signature (Radiomic, Molecular and clinical signature).

• RAD-Score associated with progression-free survival (PFS). To avoid biases, NETSPARE built from retrospective multicenter study including independent cohorts is a training cohort. The validation cohort to test the performance of the radiomic signature corresponds to the UK cohort of SIOP II Program (n=220). This division allows external validation.

Training Cohort: the median PFS is estimated to be 5 years in literature. This corresponds to a 3.5-year PFS of 62%. With 370 patients and a median follow-up of 3.5 years, we expect at least 120 events. According to the rule of thumb recommending 10 events per variable of interest, a multiparametric model combining 10 variables may be trained. An alternative modeling strategy based on a penalized approach will be used to study the association between imaging parameters and PFS. As imaging biomarker data far exceeds sample size and are intercorrelated, penalized methods (previously used in different situations with spatial data) will be used for the regularization and selection of variables, by encouraging the grouping effect and select explanatory variables.

Validation cohort: 66 events are expected. With this number of events, it is possible to detect with 80% power a hazard ratio of 0.5 between the low- and high-risk group (Logrank test two-sided 0.05).

Demographic data: Continuous variables will be summarized by cohort using median, minimum, maximum and number of available observations. Qualitative variables will be summarized using counts, percentages and number of missing data. Comparison between cohort will be performed using the chi-square test (or Fisher exact test if applicable) for qualitative variables and the Mann-Whitney test for quantitative variables.

• Radiomics signature development and validation: To study the association between imaging features and PFS, we will use an alternative modeling strategy based on the penalized Cox regression model. Elastic Net method encourages grouping effect and selects explanatory variables. We will perform a 10-fold cross-validation to select the best penalty parameter lambda. The mixing parameter α, other parameter of the Elastic net method, will be set to a default value of 0.5. The regression coefficients associated with each state of the different variables will be used to determine the RAD score. As the outcome is time-dependent, the risk score will be dichotomized using a time-dependent ROC curve to obtain a classifier (risk groups: poor vs intermediate vs good prognosis). The optimal threshold has been determined to obtain at 3 years a sensitivity of at least 80% and the best specificity.

An interval validation will be performed using a leave k-out cross-validation. We will then apply the RAD-score signature to the validation cohort. Performances of the score will be evaluated using different criteria: monotonicity of the prognostic stratification assessed using Kaplan-Meier survival curves and Hazard Ratio (HR) estimation, discrimination evaluated by Harrell's C-Index and the D statistics, quality of the model and the proportion of explained variation estimated by the Akaike Information Criterion (AIC) and the RD² statistic, respectively. The ability of prognostic score to identify patients at high risk of relapse during the first 3 years will be evaluated by determining the sensitivity, specificity and predictive value using a time-dependent ROC curve.

• RADClin-Score development and validation: To incorporate clinical and imaging informations, we will combine both low-dimensional clinical and high-dimensional imaging data in a global prediction model. First, we will completely ignore the imaging features. A Cox model will be fitted to identify clinical covariates associated with PFS. We will build a clinical linear predictor using a regression coefficient. Next, we will use Elastic Net with the clinical linear predictor in offset to identify imaging features. Finally, we will compute the RADClin-Score using a regression coefficient. We will plot a Venn diagram to compare feature selection with the RADClin-Score and the RAD-Score. We will compare the model using the different criteria previously presented.