Childhood Cancer Data Initiative Annual Symposium (Abstract Registration): Submission #67

Abstract: Whole-slide pathology images contain spatial patterns of tumor morphology that may complement clinical and molecular data, but their gigapixel scale makes these patterns difficult to search and reuse across pediatric cancer cohorts. We asked whether high-performance computing could transform CCDI whole-slide images at cohort scale into a persistent, model-agnostic vector database while preserving traceability to the originating patient, slide, and tissue region.
We developed a decoupled three-stage workflow for patch generation, foundation-model inference, and sharded vector-database ingestion on the Oak Ridge Leadership Computing Facility-Frontier exascale system. The database schema is independent of a specific embedding model and records model version, vector collection, spatial coordinates, quality measures, and linked clinical-pathology metadata. We demonstrated this design using three state-of-the-art foundational models HIPT, H-Optimus-0, and Virchow2.
The workflow processed 4,185 H&E whole-slide images from the CCDI Molecular Characterization Initiative, representing 4,054 patients and 19,603 DICOM objects. It generated 414 million image patches, including 170 million tissue-containing regions, and persisted their representations for query and reuse.
The resulting database supports combined metadata filtering and vector similarity queries, visualization, quality review, and reuse of precomputed representations without repeating extraction from raw gigapixel images. By using HPC to create a model-agnostic, metadata-aware vector database at CCDI scale, this work converts static pathology images into reusable data objects and expands the infrastructure available for future multimodal pediatric cancer research.