NCI Data Jamboree (Project Abstract Submission): Submission #54

Presenter Information
---------------------
First Name: Mayanka
Middle Initial: {Empty}
Last Name: Chandra Shekar
Degree(s): Ph.D.
Position/Title/Career Status: Research Scientist
Organization: Oak Ridge National Laboratory
Organization Address:
Knoxville

Email: chandrashekm@ornl.gov

Additional Authors
------------------
List of Additional Authors:
- First Name: Xi
  Last Name: Zhang
  Post-nominal letters: Ph.D.
  Affiliation: Oak Ridge National Laboratory
- First Name: Ankita
  Last Name: Paul
  Post-nominal letters: Ph.D.
  Affiliation: Oak Ridge National Laboratory
- First Name: Ethan
  Last Name: Seefried
  Post-nominal letters: Ph.D.
  Affiliation: Oak Ridge National Laboratory


Abstract Information
--------------------
Abstract Category: Employing statistical, computational, and informatics tools, algorithms, and methods to integrate or analyze data
Abstract Keywords: whole-slide imaging; foundation-model embeddings; 
Abstract Title: Hierarchical, Spatially Aware Whole-Slide Image Classification for Pediatric Cancer Using RADIANCE
Abstract:
Pediatric cancers are rare and morphologically heterogeneous, making it difficult to develop robust whole-slide image classification models. RADIANCE is a pathology data project that creates reusable foundation-model embeddings and links them to patch coordinates, slide and specimen identifiers, model provenance, and available clinical metadata.
During the 3-days, we will evaluate whether embeddings contain sufficient information for pediatric cancer classification. Using whole-slide images from the CCDI MCI, we will select a cohort with suitable diagnostic labels and train a model to classify cancer or histologic subtypes. We will first establish a baseline that treats patch embeddings as an unordered collection. We will then evaluate a hierarchical approach that incorporates patch locations and relationships among neighboring tissue regions. The primary technical question is whether spatially organized aggregation improves classification compared with conventional non-spatial aggregation.
This project is relevant to the broader cancer data community because generating foundation-model embeddings from whole-slide images is computationally expensive. Evaluating reusable embeddings for a concrete downstream task will help determine whether they can support multiple studies without repeatedly processing the source images. We also intend to share the resulting vector database as a resource for the community to use in whole-slide image downstream tasks. The workflow may also provide a template for classification, cohort discovery, case retrieval, and other pathology applications.
The work will require expertise in computational pathology, pediatric cancer, machine learning, and data engineering. Expected tools include Python, PyTorch, pathology foundation-model embeddings, the Milvus vector database for embedding storage and retrieval, and GPU-enabled computing on OLCF’s Frontier supercomputing platform for model training and evaluation. We have assembled a ready-to-go team from Oak Ridge National Laboratory.  
Mayanka ChandraShekar (in-person), Ethan Seefried (in-person), Xi Zhang (online), Ankita Paul (online)