Image informatics tools for curation and prognostic modeling in pediatric brain tumors

ABSTRACT: A significant challenge for ~4,000 new cases of childhood brain tumors every year is to minimize the risk of overtreatment or undertreatment, which is responsible for the wide disparity in patient outcomes. Despite revised molecularly informed risk-stratification, prognosis for many children with brain cancer remains poor with morbid long-term sequelae associated with aggressive chemoradiation. Given the wide availability of MRI within the clinical workflow, there is an opportunity for developing reliable, complementary image-based prognostic companion diagnostic tools for pediatric brain tumors and thus provide critically needed and clinically actionable information for (i) identifying high-risk cases who are most likely to receive added benefit from adjuvant & concomitant therapy, while (ii) enabling therapy de-escalation in low/standard-risk cases. Image-based companion prognostic models using machine-learning (ML) and deep learning (DL) have shown significant promise in adult tumors, including in brain tumors. However, biological differences and considerably limited data in pediatric brain tumors poses challenges in adopting similar ML/DL pipelines for prognostic modeling. These include: (1) ensuring “quality-controlled” cohorts that account for reduced tissue contrast, noise, and resolution issues in pediatric scans; (2) lack of expert-vetted, deeply annotated pediatric brain tumor scans; and (3) paucity of radiomics descriptors (computerized extraction of sub-visual information from routine imaging) designed to capture unique tumor manifestation in pediatric brain tumors while accounting for brain development. In this U01 project, we propose to develop, validate, disseminate the first unified, community driven pediatric brain tumor image informatics (PBTI2) toolkit, which will comprise three modules: (a) CuPed, a synergistic cohort curation tool which will allow for efficiently triaging imaging scans to meet user-specified bounds on image quality, as well as intelligently account for batch effects; (b) SegPed, an interactive human-in-the-loop segmentation tool for creating deeply annotated pediatric MRIs, and (c) RaPed, a suite of specialized radiomics descriptors that account for age and location specific morphometric differences in the growing brain structure, unique to pediatric brain tumors. Leveraging our access to the Children’s Brain Tumor Network (CBTN), the validation of PBTI2 will take place within two specific use-cases: (1) creating the largest repository of expert vetted, deeply annotated pediatric brain tumor datasets of 3000+ CBTN MRI scans, and (b) building and validating an image-based companion-prognostic model for survival stratification in medulloblastoma tumors via CBTN and a completed clinical trial cohort (N>500). With integration of PBTI2 modules into NCI platforms such as IDC, 3D Slicer, FeTS, and public release of expert-vetted segmentations and radiomic features, PBTI2 will have far-reaching implications in future diagnostic/prognostic models for improving outcomes in this underserved population.