Abinand Rejimon
BME PhD Proposal Presentation

Date: 2025-07-10
Time: 1:00 PM - 2:30 PM
Location / Meeting Link: C3018 at Winship Cancer Institute on Emory University Clifton Campus or Zoom (https://zoom.us/j/5582282841?omn=95527174614)

Committee Members:
Hyunsuk Shim, PhD (Advisor); Lee A. Cooper PhD (Co-advisor); Candace Fleischer, PhD; Brent Weinberg, MD/PhD; John Oshinski, PhD; Bree Eaton, MD;

Title: An Integrated Approach for Optimizing Spectroscopy-Guided Radiotherapy Strategies in Pediatric High-Grade Glioma

Abstract:
Pediatric high-grade glioma (pHGG) is the deadliest childhood brain tumor, with five-year survival below 20 percent because standard radiotherapy (RT) planning, adapted from adult imaging rules, often overlooks non-enhancing or infiltrative margins and exposes healthy brain to unnecessary radiation. Whole-brain spectroscopic MRI (sMRI) maps endogenous metabolites such as elevated choline (Cho) and reduced N-acetylaspartate (NAA) that reveal tumor spread earlier and more specifically than contrast MRI. Routine clinical adoption remains limited by two issues: (1) prospective data are still needed to verify that metabolite maps accurately reflect pediatric tumor biology and recurrence, and (2) sMRI artifacts require time-consuming expert review before RT planning. The long-term goal of this study is to increase the effectiveness of RT in pHGG, improving survival outcomes and reducing recurrence rates. The first objective of the project will utilize sMRI to guide stereotactic biopsies in pHGG patients. Increasing Cho/NAA ratios will be used to delineate biopsy targets to correlate metabolite thresholds with histology and define infiltrative cut-offs. This aim will also use longitudinal pHGG data to characterize and visualize changes in metabolites in areas surrounding RT dose. This information will be utilized to create high-risk atlases. Pilot biopsies have already separated active tumors from radiation necrosis, and early tracking shows metabolite abnormalities appear before standard MRI changes, demonstrating feasibility. The second objective of this project will automate sMRI artifact filtration. A supervised convolutional classifier will be trained on expert-labeled tumor spectra and, if necessary, refined through active-learning loops. Additionally, I will aim to create a self-supervised model which will pretrain the same encoder on a larger pool of unlabeled spectra using denoising autoencoding and masked-region reconstruction before fine-tuning on the labeled subset. Performance will be benchmarked against the supervised classifier and previous models developed in our lab. Through advancements in treatment accuracy and the ability to characterize recurrence risk, these objectives will seek to improve the management of pHGG.