Developing an AI-Based Prognostic Model for Ewing Sarcoma Using Digital Pathology Images

About This Grant

Project Summary Ewing sarcoma is a rare and aggressive pediatric bone and soft-tissue cancer. Despite advances in multimodal therapy, approximately 40% of patients die from their disease. Current risk stratification relies on broad clinical factors such as age, tumor size, and metastatic status, which fail to capture underlying histologic and biological heterogeneity. No validated biomarkers or computational models currently guide therapy, leaving clinicians unable to tailor treatment intensity to individual risk. Recent advances in artificial intelligence (AI) and digital pathology now make it possible to extract quantitative prognostic information directly from hematoxylin and eosin (H&E) whole-slide images (WSIs). Pathology foundation models—large neural networks pretrained on millions of pathology images—enable robust feature extraction even in rare diseases with limited sample sizes. Through the Children’s Oncology Group (COG), we have assembled the largest Ewing sarcoma digital pathology dataset worldwide (>900 cases with outcomes), and the Pediatric Cancer Data Commons (PCDC) provides an independent validation cohort (~100 cases). Together, these resources create an unprecedented opportunity to explore AI-based risk prediction in this rare pediatric cancer. This exploratory R21 aims to establish the feasibility and analytic rigor of histology-based prognostic modeling in Ewing sarcoma. Aim 1 will develop a prognostic model using COG WSIs to predict patient survival by leveraging foundation-model embeddings and quantitative, pathologist-defined histologic features to identify reproducible image biomarkers of outcome. Aim 2 will independently validate the locked model in the PCDC cohort without retraining. All analyses will follow TRIPOD-AI reporting guidelines to ensure transparency and reproducibility. A distinctive innovation of this study is the planned fusion of AI-derived image embeddings with pathologist-defined features—an approach not previously applied in pediatric oncology. This integration combines the sensitivity of deep learning with the interpretability of human-recognized morphology, providing a pathway toward clinically explainable AI models. If successful, this research will establish the feasibility of histology-based prognostic modeling in Ewing sarcoma and lay the groundwork for future, larger-scale studies aimed at developing validated, AI- driven biomarkers for precision risk stratification in pediatric cancers. The project brings together experts in pediatric oncology, computational pathology, and biostatistics with direct access to COG and PCDC resources, ensuring technical feasibility and translational relevance. The methods developed here will also be broadly applicable to other rare pediatric and adolescent cancers, advancing the role of AI-driven digital pathology in precision oncology.