NCI Office of Data Sharing (ODS) Data Jamboree (Abstract Submissions): Submission #9

Submission information
Submission Number: 9
Submission ID: 144817
Submission UUID: 779c98c9-5f63-4322-84a3-f9a52cbd4f31

Created: Mon, 06/16/2025 - 14:58
Completed: Mon, 06/16/2025 - 14:58
Changed: Mon, 06/16/2025 - 14:58

Remote IP address: 10.208.28.250
Submitted by: Anonymous
Language: English

Is draft: No
First Name Rakesh
Middle Initial
Last Name Khanna
Degree(s) B.A.
Position/Title/Career Status
Organization CBIIT Computational Genomics and Bioinformatics Branch
Organization Address Rockville, MD
Email rakesh.khanna@nih.gov
Other (Please Specify)
Abstract Category Employment of statistical methods or existing computational, mathematical, or informatics tools
Abstract Keywords Deep Learning, Foundation Model, Whole Slide Images, Brain, Feature Extraction
Abstract Title Extracting Deep Learning Features from Childhood Brain Tumor Whole Slide Images for Multi-Modal Analysis
Abstract Summary ### Background:
Whole slide images (WSIs) contain rich morphological information used to understand childhood brain tumors. However, accessing and processing this data requires significant computational overhead including specialized hardware and expertise.

### Methods
This project aims to faciliate a number of downstream analyses utilizing histopathological features by generating pre-computed embeddings from WSIs of 433 childhood brain tumor patients from the Childhood Cancer Data Initative-Molecular Characterization Initiative (CCDI-MCI), a heterogenous collection of CNS tumors including gliomas, astrocytomas, and medulloblastomas.

Using **TRIDENT**, a toolkit for large-scale whole-slide image processing we will generate both patch-level (UNI2-h, CONCHv1.5, and Prov-Gigapath) and slide-level representations (Threads, Titan and CHEIF).

### Purpose
By providing pre-extracted features, we enable scientists to immediately incorporate histopathological information into their own analyses. We will demonstrate the potential utility of these feature sets through proof-of-concept downstream analyses according to available clinical and omics data received at the jamboree.

All extracted features will be publicly available enabling the broader pediatric cancer community to leverage imaging data in their investigations without the typical computational overhead.
Upload Abstract