FY27 Decoding Histopathology Images for Prognostics (Decoding Spatial Cellular Architecture)

ABSTRACT: Pediatric brain tumors are the most common solid malignancy in children and the leading cause of cancer-related death in the pediatric population. Despite advances in surgery, radiation, and chemotherapy, prognosis varies widely even within the same histologic or molecular subtype. Current risk stratification methods, which rely on histopathological classification and limited molecular biomarkers, fail to capture the cellular heterogeneity and spatial organization of the tumor microenvironment (TME). These complex characteristics are increasingly recognized as critical determinants of disease progression and therapy resistance. Recent advances in multiplex imaging and spatial transcriptomics have revealed prognostic cell types and spatial patterns in various cancers. However, these technologies are expensive, tissue-destructive, and impractical for large retrospective studies that rely on formalin-fixed, paraffin-embedded archival material. In contrast, hematoxylin and eosin (H&E)-stained slides are routinely collected, inexpensive, and available across decades of archival material. These slides contain rich morphological and spatial information, yet their potential for large- scale quantitative TME profiling in pediatric neuro-oncology remains underexplored. Existing artificial intelligent (AI)-based approaches face key limitations for pediatric application: they often require labor-intensive pixel-level manual annotations, are trained predominantly on adult tumor morphology, and rarely resolve fine-grained malignant cell states alongside diverse non-malignant cell types. The overall goal of this project is to develop and apply a deep learning-based pathology framework that bridges histopathology and spatial tumor biology for prognostic biomarker discovery in pediatric brain tumors. Aim 1 will develop a weakly supervised, multi-scale deep learning framework to infer tumor cell states and TME composition directly from H&E slides by integrating histopathologic features with spatial transcriptomic references from pediatric brain tumors. The model will leverage spot-level molecular annotations and eliminate the need for pixel-level labels. It employs self- supervised and attention-based architectures to capture both fine-grained malignant cell heterogeneity and the diversity of non-malignant cell types in the TME. This pediatric-specific model will address limitations of manual review and adult-tumor focused AI approaches, enabling accurate, interpretable, and scalable TME profiling from histopathological images. Aim 2 will apply this framework to large-scale pediatric brain tumor pathology datasets to reconstruct spatial cellular architectures and quantify their association with patient outcomes. Prognostic models will be built to identify cell-type- and spatial pattern-based biomarkers predictive of survival and recurrence, validated across independent multi-institutional cohorts. This project will establish a scalable, interpretable, and non-destructive approach for high-throughput TME profiling in pediatric brain tumors, filling a critical methodological gap between routine histopathology and high-resolution spatial omics. The resulting framework will not only advance prognostic risk stratification but also uncover spatial biomarkers to inform the design of future clinical trials and targeted therapies, ultimately improving outcomes for children with brain tumors.