A multi-institutional pediatric dataset of clinical radiology MRIs by the Children's Brain Tumor Network

Pediatric brain and spinal cancers remain the leading cause of cancer-related death in children. Advancements in clinical decision-support in pediatric neuro-oncology utilizing the wealth of radiology imaging data collected through standard care, however, has significantly lagged other domains. Such data is ripe for use with predictive analytics such as artificial intelligence (AI) methods, which require large datasets. To address this unmet need, we provide a multi-institutional, large-scale pediatric dataset of 23,101 multi-parametric MRI exams acquired through routine care for 1,526 brain tumor patients, as part of the Children's Brain Tumor Network. This includes longitudinal MRIs across various cancer diagnoses, with associated patient-level clinical information, digital pathology slides, as well as tissue genotype and omics data. To facilitate downstream analysis, treatment-naïve images for 370 subjects were processed and released through the NCI Childhood Cancer Data Initiative via the Cancer Data Service. Through ongoing efforts to continuously build these imaging repositories, our aim is to accelerate discovery and translational AI models with real-world data, to ultimately empower precision medicine for children.

Automated Tumor Segmentation and Brain Tissue Extraction from Multiparametric MRI of Pediatric Brain Tumors: A Multi-Institutional Study

Background

Brain tumors are the most common solid tumors and the leading cause of cancer-related death among all childhood cancers. Tumor segmentation is essential in surgical and treatment planning, and response assessment and monitoring. However, manual segmentation is time-consuming and has high interoperator variability. We present a multi-institutional deep learning-based method for automated brain extraction and segmentation of pediatric brain tumors based on multi-parametric MRI scans.

Methods

Multi-parametric scans (T1w, T1w-CE, T2, and T2-FLAIR) of 244 pediatric patients (n=215 internal and n=29 external cohorts) with de novo brain tumors, including a variety of tumor subtypes, were preprocessed and manually segmented to identify the brain tissue and tumor subregions into four tumor subregions, i.e., enhancing tumor (ET), non-enhancing tumor (NET), cystic components (CC), and peritumoral edema (ED). The internal cohort was split into training (n=151), validation (n=43), and withheld internal test (n=21) subsets. DeepMedic, a three-dimensional convolutional neural network, was trained and the model parameters were tuned. Finally, the network was evaluated on the withheld internal and external test cohorts.

Results

Dice similarity score (median±SD) was 0.91±0.10/0.88±0.16 for the whole tumor, 0.73±0.27/0.84±0.29 for ET, 0.79±19/0.74±0.27 for union of all non-enhancing components (i.e., NET, CC, ED), and 0.98±0.02 for brain tissue in both internal/external test sets.

Conclusions

Our proposed automated brain extraction and tumor subregion segmentation models demonstrated accurate performance on segmentation of the brain tissue and whole tumor regions in pediatric brain tumors and can facilitate detection of abnormal regions for further clinical measurements.

Unsupervised machine learning using K-means identifies radiomic subgroups of pediatric low-grade gliomas that correlate with key molecular markers

INTRODUCTION

Despite advancements in molecular and histopathologic characterization of pediatric low-grade gliomas (pLGGs), there remains significant phenotypic heterogeneity among tumors with similar categorizations. We hypothesized that an unsupervised machine learning approach based on radiomic features may reveal distinct pLGG imaging subtypes.

METHODS

Multi-parametric MR images (T1 pre- and post-contrast, T2, and T2 FLAIR) from 157 patients with pLGGs were collected and 881 quantitative radiomic features were extracted from tumorous region. Clustering was performed using K-means after applying principal component analysis (PCA) for feature dimensionality reduction. Molecular and demographic data was obtained from the PedCBioportal and compared between imaging subtypes.

RESULTS

K-means identified three distinct imaging-based subtypes. Subtypes differed in mutational frequencies of BRAF (p < 0.05) as well as the gene expression of BRAF (p<0.05). It was also found that age (p < 0.05), tumor location (p < 0.01), and tumor histology (p < 0.0001) differed significantly between the imaging subtypes.

CONCLUSION

In this exploratory work, it was found that clustering of pLGGs based on radiomic features identifies distinct, imaging-based subtypes that correlate with important molecular markers and demographic details. This finding supports the notion that incorporation of radiomic data could augment our ability to better characterize pLGGs.

Automated Tumor Segmentation and Brain Tissue Extraction from Multiparametric MRI of Pediatric Brain Tumors: A Multi-Institutional Study

Background

Brain tumors are the most common solid tumors and the leading cause of cancer-related death among all childhood cancers. Tumor segmentation is essential in surgical and treatment planning, and response assessment and monitoring. However, manual segmentation is time-consuming and has high interoperator variability. We present a multi-institutional deep learning-based method for automated brain extraction and segmentation of pediatric brain tumors based on multi-parametric MRI scans.

Methods

Multi-parametric scans (T1w, T1w-CE, T2, and T2-FLAIR) of 244 pediatric patients (n=215 internal and n=29 external cohorts) with de novo brain tumors, including a variety of tumor subtypes, were preprocessed and manually segmented to identify the brain tissue and tumor subregions into four tumor subregions, i.e., enhancing tumor (ET), non-enhancing tumor (NET), cystic components (CC), and peritumoral edema (ED). The internal cohort was split into training (n=151), validation (n=43), and withheld internal test (n=21) subsets. DeepMedic, a three-dimensional convolutional neural network, was trained and the model parameters were tuned. Finally, the network was evaluated on the withheld internal and external test cohorts.

Results

Dice similarity score (median±SD) was 0.91±0.10/0.88±0.16 for the whole tumor, 0.73±0.27/0.84±0.29 for ET, 0.79±19/0.74±0.27 for union of all non-enhancing components (i.e., NET, CC, ED), and 0.98±0.02 for brain tissue in both internal/external test sets.

Conclusions

Our proposed automated brain extraction and tumor subregion segmentation models demonstrated accurate performance on segmentation of the brain tissue and whole tumor regions in pediatric brain tumors and can facilitate detection of abnormal regions for further clinical measurements.

Key Points

We proposed automated tumor segmentation and brain extraction on pediatric MRI.The volumetric measurements using our models agree with ground truth segmentations.

Importance of the Study

The current response assessment in pediatric brain tumors (PBTs) is currently based on bidirectional or 2D measurements, which underestimate the size of non-spherical and complex PBTs in children compared to volumetric or 3D methods. There is a need for development of automated methods to reduce manual burden and intra- and inter-rater variability to segment tumor subregions and assess volumetric changes. Most currently available automated segmentation tools are developed on adult brain tumors, and therefore, do not generalize well to PBTs that have different radiological appearances. To address this, we propose a deep learning (DL) auto-segmentation method that shows promising results in PBTs, collected from a publicly available large-scale imaging dataset (Children's Brain Tumor Network; CBTN) that comprises multi-parametric MRI scans of multiple PBT types acquired across multiple institutions on different scanners and protocols. As a complementary to tumor segmentation, we propose an automated DL model for brain tissue extraction.

NIMG-102. RAPNO-DEFINED SEGMENTATION AND VOLUMETRIC ASSESSMENT OF PEDIATRIC BRAIN TUMORS ON MULTI-PARAMETRIC MRI SCANS USING DEEP LEARNING; A ROBUST TOOL WITH POTENTIAL APPLICATION IN TUMOR RESPONSE ASSESSMENT

Volumetric measurements of whole tumor and its components on MRI scans, facilitated by automatic segmentation tools, are essential to reduce inter-observer variability in monitoring tumor progression and response assessment for pediatric brain tumors. Here, we present a fully automatic segmentation model based on deep learning that reliably delineates the tumor components recommended by the Response Assessment in Pediatric Neuro-Oncology (RAPNO) working group for evaluation of treatment response. Multi-parametric MRI (mpMRI) scans (T1-pre, T1-post, T2, and T2-FLAIR), acquired on multiple MRI scanners with different field strengths and vendors, for a cohort of 218 pediatric patients with a variety of histologically confirmed brain tumor subtypes were collected. The mpMRI scans were co-registered and manually segmented by experienced neuroradiologists in consensus to identify the tumor subregions including the enhancing tumor (ET), non-enhancing tumor (NET), cystic components (CC), and peritumoral edema (ED) regions. A convolutional neural network model based on DeepMedic architecture was trained using mpMRI scans as the inputs for segmentation of the whole tumor and subregions. The trained model showed excellent performance in segmentation of the whole tumor, as suggested by median dice of 0.90/0.85 for validation (n = 44)/independent test (n = 22) sets. ET and non-enhancing components (union of NET, CC, and ED) were segmented with median dice scores of 0.78/0.84 and 0.76/0.74 for validation/test sets, respectively. The automated and manual segmentations demonstrated strong agreement in estimating VASARI (Visually AcceSAble Rembrandt Images) MRI features with Pearson’s correlation coefficient R &gt; 0.75 (p &lt; 0.0001) for ET, NET, CC, and ED components. Our proposed automated segmentation method developed based on MRI scans acquired with different protocols, equipment, and from a variety of brain tumor subtypes, shows potential application for reliable and generalizable volumetric measurements which can be used for treatment response assessment in clinical trials.

NIMG-74. RADIOIMMUNOMIC SIGNATURES IN PEDIATRIC LOW-GRADE GLIOMA BASED ON MULTIPARAMETRIC MRI SCANS

Understanding the immune microenvironment in pediatric low-grade glioma (pLGG) patients may help in identification of the patients who benefit from anti-tumor immunotherapies. However, surgical resection is not feasible for many pLGG tumors in certain anatomical locations. Therefore, developing non-invasive tools that characterize the tumor microenvironment prior to therapeutic interventions could contribute to stratification and enrollment of the patients into relevant clinical trials. In this work, we derived radiomic signatures of immune profiles (radioimmunomics) based on machine learning (ML) analysis of readily available conventional MRI scans. Transcriptomic data for a cohort of 197 subjects was retrospectively collected from Open Pediatric Brain Tumor Atlas (OpenPBTA). The patients were categorized into three groups (Group1-3) based on their immunological profiles using consensus clustering algorithm. This analysis revealed greater immune cell infiltration in non-BRAF mutated pLGGs. Group1 showed more enrichment in M1 macrophages, and microenvironment and immune scores compared to Group2 and Group3. Elevated tumor inflammation score (TIS), as a predictor of clinical response to anti-PD-1 blockade, was observed in Group1 compared to Group2 (p= 1.4e-7) and Group3 (p= 0.0054). Radiomic features, including volumetric, morphologic, histogram, and texture descriptors, were extracted from the segmented tumor regions on multiparametric MRI (mpMRI) scans of 71 (of 197) patients. Multivariate ML models were trained to predict the three immunological groups based on radiomic features using cross-validated random forest classifier along with recursive feature elimination, which yielded AUC of 0.72 for this multi-class classification problem. Our findings indicate the presence of distinct immunological groups in pLGG tumors, with possibly more favorable response to immunotherapies in Group1 tumors. Furthermore, we developed radioimmunomic signatures based on pre-operative conventional mpMRI that can potentially stratify the patients based on their immune tumor microenvironment. Based on these initial promising results, we are exploring additional features to increase the accuracy of radioimmunomics model.

NIMG-87. CHARACTERIZING IMMUNE PROFILES OF PEDIATRIC MEDULLOBLASTOMA AND THEIR RADIOLOGICAL CORRELATES

Recent studies have shown preliminary evidence for differentiation of the tumor microenvironment (TME) and immune landscape between molecularly-defined medulloblastoma (MB) subtypes. Identifying radiological correlates of these TME patterns could establish a non-invasive method of immune profile characterization for guiding patient-centered therapies. Here, we examine immune profiles between MB subtypes using data from Open Pediatric Brain Tumor Atlas (OpenPBTA), and their relationship to tumor measurements from pre-operative MRIs. We identified a retrospective cohort of 94 pediatric MB patients with available molecular subtyping and immune profiles (36 cell types) from bulk gene expression data. A random forest analysis was used to classify the four MB subtypes based on immune profiles. Four cell types had high impact on classification performance: plasmacytoid dendritic cells (PDC; 25.8% accuracy decrease when randomized), hematopoietic stem cells (HSC; 21.9%), plasma B cells (20.3%), and cancer associated fibroblasts (18.8%). Pairwise comparisons revealed SHH and WNT tumors had significantly higher numbers of fibroblasts and HSCs compared to Group3/Group4. We also found novel evidence for significantly lower amounts of plasma B cells in the SHH group, and high PDC levels in Group4, followed by Group3, and low PDC in SHH/WNT. Multi-parametric MRI scans for 39 patients were used to segment tumor volumes. Overall tumor volume was significantly correlated with composite stroma scores (R = 0.34, p = 0.036). Additionally, patients with higher volumes of gadolinium contrast-enhancing compared to non-enhancing components had higher immune (R = 0.42, p = 0.009) and microenvironment (summed immune and stromal cell types; R = 0.44, p = 0.006) scores, regardless of their molecular subtype. Together, our results demonstrate: (1) the use of rich immune profiles for differentiating molecular subtypes of MB and their unique TME characterization; and (2) initial evidence for radiological correlates of these profiles based on pre-operative imaging collected through standard practices.

NIMG-62. RADIOMIC-BASED PROGRESSION-FREE SURVIVAL STRATIFICATION OF PEDIATRIC LOW-GRADE GLIOMA IS ASSOCIATED WITH MOLECULAR ALTERATIONS

Pediatric low-grade glioma (pLGG) encompasses a variety of tumor subtypes with heterogeneous treatment response and relatively long progression-free survival (PFS). Radiomics may serve as a non-invasive and in-vivo tool for early prediction of PFS as a surrogate marker for treatment response and to objectively gauge the efficacy of novel treatment strategies. Here, we present a multivariate model based on radiomic features and clinical variables for risk stratification of pLGGs in terms of PFS and seek associations of the predicted risk groups and mutations in key molecular markers using data from PedCBioportal. Pre-operative multi-parametric MRI scans (T1-pre, T1-post, T2, T2-FLAIR) of 129 patients with newly diagnosed pLGG (median age, 7.76, range, 0.35-19.58 years; median PFS, 28.5, range, 1.1-124.8 months) were collected and quantitative radiomic features (n = 881) were extracted. A multivariate Cox proportional hazard’s (Cox-PH) regression model was fitted based on clinical (age, sex, and extent of tumor resection) and radiomic variables using 4-fold cross-validation. A subset of radiomic features (n = 27) that were most predictive of PFS was selected by applying Elastic Net regularization penalty during Cox-PH model fitting. High-, medium- and low-risk groups were determined based on model predictions. Cox-PH modeling showed excellent performance for prediction of PFS as suggested by the concordance index of 0.78. Radiogenomic assessment (data available in 94/129 patients) showed more enrichment of mutations in NF1 and RB1 genes in the high-risk group, as compared to the low- and medium-risk groups. We showed the potential value of radiomics in providing upfront prediction of PFS, which may further be used as an added treatment arm for early assessment of treatment response of the pLGG patients enrolled in the clinical trials. In the next step of this work, we will expand the cohort and cross-validate these results in an external cohort.

IMG-15. Radiomic Profiling of Pediatric Low-Grade Glioma Improves Risk Stratification Beyond Clinical Measures

PURPOSE: Treatment response is heterogeneous among patients with pediatric low-grade glioma (pLGG), the most frequent childhood brain tumor. Upfront prediction of progression-free survival (PFS) may facilitate more personalized treatment planning and improve outcomes for the pLGG patients. In this work, we explored the additive value of radiomics to clinical measures for prediction of PFS in pLGGs. We further sought associations between the derived risk groups and underlying alterations in key genomic and transcriptomic variables. METHODS: Quantitative radiomic features were extracted from pre-operative multi-parametric MRI scans (T1, T1-post, T2, T2-FLAIR) of 96 patients with newly diagnosed pLGG (median age, 8.59, range, 0.35-18.87 years; median PFS, 25.23, range, 3.03-124.83 months). Multivariate Cox proportional hazard’s (Cox-PH) regression models were fitted using 5-fold cross-validation on a training cohort of 68 subjects and tested on 28 patients. Three models were generated using (1) only clinical variables (age, sex, and extent of tumor resection), (2) radiomic features, and (3) clinical and radiomic variables. The dimensionality of radiomic features in Cox-PH models was reduced by applying Elastic Net regularization penalty to identify a subset of variables that are most predictive of PFS. The patients were then stratified into three groups of high, medium, and low-risk based on model predictions. RESULTS: Cox-PH modeling resulted in a concordance index (c-index) of 0.55 for clinical data, 0.65 for radiomics, and 0.73 for a combination of clinical and radiomic variables, highlighting the additive value of radiomics to the readily available clinical information in prediction of PFS. Radiogenomic assessments revealed significant differences in expression of BRAF, NF1, TSC1, ALK (p&lt;0.01), and RB1 (p&lt;0.05) genes in the high-risk group, compared to the medium and low-risk groups. CONCLUSION: Our results demonstrate the value of integrating radiomics with clinical measures to improve risk assessment of patients with pLGG through improved pretreatment prediction of PFS.

Radiomics and radiogenomics in pediatric neuro-oncology: A review

The current era of advanced computing has allowed for the development and implementation of the field of radiomics. In pediatric neuro-oncology, radiomics has been applied in determination of tumor histology, identification of disseminated disease, prognostication, and molecular classification of tumors (ie, radiogenomics). The field also comes with many challenges, such as limitations in study sample sizes, class imbalance, generalizability of the methods, and data harmonization across imaging centers. The aim of this review paper is twofold: first, to summarize existing literature in radiomics of pediatric neuro-oncology; second, to distill the themes and challenges of the field and discuss future directions in both a clinical and technical context.