RADT-26. IMPROVING RADIATION TARGET VOLUME DEFINITION FOR GLIOBLASTOMA USING PREDICTIONS OF TUMOR RECURRENCE FROM MACHINE LEARNING AND PRE-RADIOTHERAPY ADVANCED MRI

• DOI: 10.1093/neuonc/noad179.0215
• Publication Date: 2023-11-11T23:35:15Z
• Authors: Nate Tran; Tracy Luks; Jacob Ellison; Devika Nair; Yan Li; Angela Jakary; Bo Liu; Oluwaseun Adegbite; Javier Villanueva...; Nicholas Butowski; Jennifer Clarke; Susan Chang; Steve Braunstein; Hui Lin; Olivier Morin; Janine Lupo
• Keywords: 0301 basic medicine; 03 medical and health sciences; 3. Good health

Abstract

INTRODUCTION Standard-of-care (SOC) radiation therapy (RT) planning only utilizes a fairly 1.5-2cm uniform expansion of T2-weighted-FLAIR MRI lesion to generate a clinical target volume (2cm-CTV), without considering the spatial heterogeneity and infiltrative nature of glioblastomas. This study aimed to use multi-parametric MRI with 2 artificial intelligence (AI)-based approaches to predict regions of subsequent tumor progression and compare the resulting predictions to the 2cm-CTV, with the hypothesis that applying deep learning will lead to improved detection of subclinical disease that is under-treated, and more accurately delineate areas at higher risk for progression, than the uniform 2cm-CTV.

METHODS We used normalized maps of anatomical, diffusion, and metabolic MRI post-surgical resection and before RT and chemotherapy from 72 patients newly-diagnosed with glioblastoma to train Random-Forest and deep-learning UNet models to identify voxels that later exhibit progression by either the contrast-enhancing-lesion (CEL) or T2-lesion (T2L). All models were trained/validated on 54 and tested on 18 patients after careful inter-exam alignment and normalization. Model performance was compared to the 2cm-CTV using a Progression-Coverage-Coefficient (PCC), a weighted Dice coefficient that accounts for tumor-size.

RESULTS Random-Forest models were able to predict subsequent contrast-enhancing (CE) and T2 non-enhancing (NE) progression with an average respective test AUC of 0.88 and 0.82, with 15.6%/9.6% higher performance on patients who progressed early. The 2cm-CTV treatment plan achieved the highest sensitivity (0.833 vs 0.795) but also the lowest specificity (0.879 vs 0.935), over-treating normal-appearing-brain. Our deep learning model outperformed the 2cm-CTV in covering the progressed lesion, with the highest specificity (0.935 vs 0.879), PCC (0.741 vs 0.734), and had comparable sensitivity (0.795) to the 2cm-CTV, while sparing more normal brain compared to the 2cm-CTV.

CONCLUSION This study demonstrates the potential benefit of using multi-parametric MRI with deep learning to assist in RT treatment planning to reduce excessive treatment of normal brain parenchyma.