Machine Learning and Radiomics Techniques for Analysis of Daily MRI in Glioblastoma Patients

PROJECT SUMMARY Glioblastoma is the most common primary brain cancer worldwide. Novel treatment strategies are urgently needed since glioblastoma is nearly universally fatal with a median overall survival of only 1.5- 2 years. A frustrating aspect of glioblastoma is that approximately half of all patients will have what looks to be tumor growth on their post-treatment MRI, termed progression. Although, half of patients with progression will turn out to have pseudoprogression, which is a not-fully understood phenomenon believed to be edema and inflammation caused by the immune system and represents a good response to treatment. In fact, patients with pseudoprogression tend to do better than the general glioblastoma population and have a median overall survival of up to 3 years. On the other hand, patients with true progression of disease (tumor growth and poor/nonresponse to treatment) tend to do worse than the general glioblastoma population and have a medial overall survival of only 10 months. The frustrating part for clinical team, and the patients themselves, is that true progression and pseudoprogression are not discernable from one another during treatment, or even on initial post-treatment imaging (1-month post-treatment). Instead, the current gold-standard to distinguish between true and pseudoprogression is to “watch and wait” – continue monitoring with serial imaging and see if the patient clinically worsens or stabilizes. Thus, there is an unmet need for techniques that reliably and accurately determine if tumor growth/progression is occurring during treatment and predict/determine which sub-type of progression (true progression or pseudoprogression) a patient has. My laboratory focuses on responding to this unmet need through a variety of methods: serial multiparametric MRI (anatomic, perfusion, diffusion, spectroscopic, etc.), quantitative MRI analysis, machine learning, and molecular research including analyzing blood samples of glioblastoma patients to look for circulating tumor cells and other molecular markers. This proposal focuses on auto-detection of tumors on MRI based on machine learning (Aim 1) and analysis of anatomic and physiologic changes (Aim 2) from daily multiparametric MRI to address this issue by creating techniques that can detect enlarging tumors during treatment and predict between true and pseudoprogression months earlier than current methods. The goal of this proposal is to develop tools that identify and monitor patients with significant anatomic and/or physiologic tumor changes much earlier than current methods, so that in the future, prompt, aggressive, and early therapy adaption can be implemented. This project will translate directly to the practice of clinical medicine and advance the field of glioblastoma treatment. Additionally, it will allow me to gain hands-on skills and expertise in machine learning, radiomics, MRI, neuroimaging, neuro-anatomy, radiation therapy, and oncology, and aid in preparing me for a career as an academic physician scientist in the field of radiation oncology.