MR Fingerprinting for Characterization and Prediction of Glioblastoma Infiltration - PROJECT SUMMARY Glioblastoma (GBM) is the most common and aggressive adult brain tumor with a median survival time of 15 months. Medical imaging is crucial in delineating GBM tumor extent for surgical and radiotherapy planning, with magnetic resonance imaging (MRI) being the gold standard due to its excellent soft tissue contrast. Unfortunately, even MRI is known to underestimate the full extent of GBM tumors, with stereotactic biopsies revealing microscopic tumor infiltration beyond visible enhancing tissue borders and into adjacent edematous regions. Although recurrence overwhelmingly occurs in this edema-rich, peritumoral region, there is no consensus regarding inclusion of peritumor within the clinical target volume. This is because no existing non-invasive methods can successfully discriminate GBM infiltration from non-malignant peritumoral edema. To address this, there have been extensive efforts to develop MRI-based artificial intelligence prediction models for GBM infiltration. Despite promising results, these infiltration models often have poor reproducibility and generalizability, which is due to MRI’s high sensitivity to differences in acquisition protocol, scanner hardware, and processing methods. As a result, GBM recurrence is inevitable and currently incurable. Magnetic resonance fingerprinting (MRF) could potentially solve this problem. MRF is a quantitative MRI acquisition framework for rapid and robust multiparametric mapping of intrinsic tissue properties. MRF maps reflect quantitative measures of physical tissue characteristics, which leads MRF to have superior sensitivity and reproducibility compared to conventional MRI. These factors make MRF an ideal imaging technique for accurate, reproducible, and generalizable radiomic modeling of GBM infiltration. The overall objective of this project is thus to develop and validate magnetic resonance fingerprinting (MRF) artificial intelligence (AI) models for prediction of infiltrated GBM peritumor. This project has two specific aims. In Aim 1, defining and discriminative MRF image features of infiltrated peritumor will be identified to train an MRF radiomic model for pre-operative GBM infiltration prediction. We will implement two technical improvements for infiltration modeling, namely transfer learning via data-driven infiltration risk priors and semi-supervised learning using unlabeled peritumor. The developed MRF radiomic model will then be applied in Aim 2 for longitudinal assessment of radiotherapy (RT) changes. GBM patients undergoing RT will be scanned with MRF and MRI across four time points. At each time point, the developed infiltration model will be applied to generate whole tumor infiltration maps and evaluate peritumoral infiltration load. Longitudinal MRF and MRI image differences between RT-treated and non-treated peritumor will be used to optimize the infiltration model developed in Aim 1 to predict future recurrence regions. This project’s success will demonstrate the feasibility of quantitative MRF radiomics for GBM infiltration prediction, paving the groundwork for targeted GBM therapy to reduce recurrence and extend survival. The technical framework developed in this project will also be highly translatable to other studies leveraging quantitative MRI image analytics to improve brain tumor care.
