Glioblastoma (GBM) is the most common primary adult brain tumor with an incidence rate of 3.2 per 100,000 people. Due to its heterogeneous genetic characteristics, GBM carries a dismal prognosis, with a median survival of only 14 months and five-year survival rates are less than 10%. The current standard of care is maximal safe surgical resection, chemoradiation, and adjuvant temozolomide. Within the natural history of GBM, there are adaptive genetic changes within the tumor that lead to treatment resistance and inevitable recurrence, leading to patient death. While a variety of treatments can be administered for tumor recurrence, there is currently no consensus on therapy for recurrent tumor as none have been proven to provide substantial survival benefit. The major limitation of the current treatment strategy is that clinicians do not have a reliable method of longitudinally assessing tumor volumes and regional genetic characteristics of the tumor during the course of treatment. Rather, clinical decision-making is based on a manual and variable two- dimensional measure of tumor burden, a surrogate of tumor volume, and genetic characterization of select molecular markers at the time of initial surgery. A tool that can automatically assess tumor volumes and regional genetic characteristics longitudinally will substantially improve evaluation of treatment efficacy, allowing for an earlier switch to alternative treatment strategies and thus, more personalized tailoring of patient care. Thus, a critical need exists for automatic methods that non-invasively evaluate treatment efficacy on a patient-to-patient basis. To address this problem, we will develop a novel solution based on deep learning that leverages structural, diffusion, and perfusion information from multi-parametric magnetic resonance imaging. At the core of our solution is a convolutional neural network; a machine learning technique that can be trained on raw image data to predict clinical outputs of interest. Firstly, we will develop a fully automatic technique for longitudinal tracking of tumor volumes. To do this, we will develop novel deep learning architectures through incorporation of state-of-the-art neural network components that can segment both whole tumor and tumor subregions (edema, non-enhancing tumor, and gadolinium contrast-enhancing tumor). To prove algorithm utility, we will automatically derive tumor volumes in a longitudinal patient cohort and correlate volumes with clinical outcomes. Secondly, we will develop a non-invasive, deep learning algorithm for evaluation of regional genetic characteristics of GBM. To train this algorithm, we will acquire imaging-localized surgical biopsies and genetic profiling of GBM patients undergoing surgery. Once trained, the algorithm can be used to non- invasively identify clonal populations and track genetic changes associated with clinical outcomes during the course of treatment. The development of these deep learning algorithms will transform physician?s capacity for clinical decision-making and dramatically improve outcomes for a devastating disease.
The proposed research will provide automatic deep-learning tools to aid clinical decision-making for glioblastoma, the most common primary adult brain tumor that carries a dismal prognosis. The major limitation of the current treatment strategy is that clinicians do not have a reliable method of longitudinally assessing tumor volumes and regional genetic characteristics of the tumor during the course of treatment. A tool that can automatically assess tumor volumes and regional genetic characteristics longitudinally will substantially improve evaluation of treatment efficacy, allowing for an earlier switch to alternative treatment strategies and thus, more personalized tailoring of patient care and enhancement of clinical outcomes.
