Machine learning enables the use of spectroscopic MRI to guide radiation therapy in patients with glioblastoma

Abstract

Glioblastoma is the most common adult primary brain tumor and is highly aggressive due to its diffusely infiltrative nature. Radiation therapy has been shown to be the best single treatment for improving prognosis but requires accurate pre-therapy imaging for proper radiation planning. Spectroscopic magnetic resonance imaging (sMRI) is an advanced imaging modality that measures specific in vivo metabolite levels within the brain and has shown to be highly sensitive and specific in the detection of proliferative pathology. Clinical application of sMRI has been extremely limited due to computational challenges in sMRI data analysis. In this work, we utilize novel machine learning architectures to develop a software framework to close the gap for clinical utilization of sMRI in radiation therapy planning. First, we develop convolutional neural network to identify and remove spectral artifacts that lead to erroneous measurement. Next, we develop an algorithm for internally normalizing sMRI volumes, enabling voxel-to-voxel comparison across subjects and allowing threshold-based techniques to be used for target delineation. Third, we create a novel unsupervised learning framework to perform accelerated spectral quantitation, reducing the computational time and power needed to utilize sMRI. Finally, we develop a web-based software framework that bridges the gap between sMRI and its clinical use and demonstrate the feasibility of using this software in a multi-site clinical study to guide a radiation boost to regions of metabolic abnormality in patients with glioblastoma.