Detection in Progress - A Multimodal Segmentation-based Approach for Predicting Glioblastoma Recurrence

Harshita Kukreja, Nate Tran, Bo Liu, Tracy L Luks, Angela Jakary, Oluwaseun Adegbite, Yan Li, Annette M Molinaro, Javier E Villanueva-Meyer, Steve E Braunstein, Hui Lin, Janine Lupo

Submission Track: Short papers presenting ongoing research or work submitted to other venues (up to 5 pages, excluding references)

Keywords: Brain Tumor Progression Segmentation, Machine Learning for Healthcare, Lesion Segmentation Metric, Multimodal Vision Transformer

TL;DR: We propose a lesion-size-aware segmentation objective for defining RT target volumes in GBM. Our method captures recurrence beyond 2 cm margin, spares normal brain, and our metric measures lesion segmentation more accurately than traditional scores.

Abstract: Radiation therapy planning for patients with glioblastoma requires defining the clinical-target-volume by delineating the tumor and including a margin of healthy tissue to account for microscopic tumor spread post radiation therapy. The current standard-of-care practice for defining the clinical-target-volume still employs an isotropic 1–2 cm expansion of the identified T2-hyperintensity lesion. As a consequence, normal-appearing brain tissue is overtreated, and it also ends up missing progression regions as it overlooks the heterogeneous infiltrative nature of these tumors. We propose incorporating anatomical, metabolic, and diffusion-weighted imaging acquired before surgical resection or between surgery and radiation therapy, using a lesion-size-aware segmentation objective to improve clinical-target-volume definition. The results in multiple metrics demonstrate better prediction of tumor progression than standard of care, as indicated by contrast enhancement and T2-hyperintensity at recurrence. Overall, our approach minimizes treatment of normal-appearing brain and captures progressed voxels beyond the 2 cm expansion.

Submission Number: 58