Glioblastoma is the most common adult primary brain tumor and is highly aggressive in its disease course. Despite advances in neurosurgical resection, radiation targeting, and chemotherapy, the prognosis remains grim with a median survival of just 15 months. The effectiveness of current radiation therapy strategies is severely limited by shortcomings in the imaging modalities used to develop treatment plans. Current radiation therapy planning is mainly based on contrast-enhanced T1-weighted MRI, which identifies high grade tumors that are immediately associated with leaky neovasculature. Although it is an excellent diagnostic tool to identify high grade from low grade tumors, it is unable to signal occult infiltration beyond the core of the tumor. Though many believe GBM to be an incurable disease, we believe we have identified a method for optimizing tumor targeting that will increase the effectiveness of radiation therapy. A significant component of the current problem in GBM therapy is the lack of treatment for non-enhancing regions that are significantly infiltrated by neoplastic glioma cells without neovascularization. This untreated population undoubtedly leads to early recurrence. The proposed study addresses an important step toward translating an advanced quantitative imaging modality that complements the conventional imaging that is capable of reliably revealing glioma- infiltrated regions for precise, personalized treatment targeting. Proton spectroscopic magnetic resonance imaging (sMRI) is an alternative modality able to identify endogenous metabolism within tissue without the need for exogenous contrast, and has been shown to identify the metabolic abnormalities associated with tumor beyond the regions identified by T1-weighted MRI. The clinical integration of sMRI in patient management has been limited due to the computational challenges of analysis of sMRI data. Two key hurdles to be overcome are the insufficiency of filters to remove image artifacts and the necessity of quantification of metabolic levels relative to a patient's baseline metabolism. As a result, sMRI processing requires skilled user intervention and many hours of computational and user time. To automate this pipeline and provide clinically useful information to oncologists, we seek to develop a software framework for the automated and expedient processing of sMRI for use in radiation therapy planning. We will use novel advances in the fields of high performance computing and deep learning, an approach to computation that has shattered benchmarks in many medical and non-medical problems. Specifically, we will develop filters for removing artifacts, algorithms for personalized diagnosis of tumor infiltration, and explore deep learning as a method to synthesize sMRI data with anatomical and clinical metrics in a fully automated fashion. Success in the proposed work will produce a ?scanner-to-clinician? platform for quantitative, expedient, and objective analysis methods to integrate sMRI into the clinical radiation therapy planning paradigm. Ultimately, we believe this additional modality in the physician's tool belt will lead to better outcomes in patients suffering from this debilitating disease.
Proton spectroscopic magnetic resonance imaging (sMRI) is a molecular imaging modality that can accurately detect abnormalities in the metabolic signatures of tumor cells, making it ideal for the detection and monitoring of highly infiltrative brain tumors like glioblastoma. Clinical adoption of sMRI has been slow due to limited tools for expedient and objective processing of sMRI data into information that oncologists can use for therapy planning. This proposal addresses these challenges using recent advances in high performance computing and machine learning to develop a software framework for using sMRI in the clinical management of patients with glioblastoma.
|Gurbani, Saumya S; Schreibmann, Eduard; Maudsley, Andrew A et al. (2018) A convolutional neural network to filter artifacts in spectroscopic MRI. Magn Reson Med 80:1765-1775|