Deep Neural Networks for Modeling Sequential Prediction Tasks with Applications in Brain Tumors
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction, presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractGliomas are malignant brain tumors that are associated with high neurological morbidity and poor outcomes. Patients diagnosed with low-grade gliomas are typically followed by a sequence of measurements of the tumor size as they visit their medical physician. To optimize the timing of therapy of patients, the effective methodologies are needed to predict the future behavior of gliomas by modeling the mechanisms that mediate the characteristic features of gliomas. Machine learning is deployed in a range of computing tasks where designing and programming explicit algorithms with good performance is difficult or infeasible, example applications include email filtering, optical character recognition (OCR), and computer vision. In the broad sweep of machine learning’s current worldly ambitions, healthcare applications seem to top the list for funding and press in the last three years. In this thesis, we demonstrate the promise of Long Short-Term Memory Neural Networks (LSTMs), and a 1-Dimensional Convolutional Neural Network (1-D CNN) to address two important clinical questions in low-grade gliomas: 1) classification and prediction of future behavior; and 2) early detection of dedifferentiation to a higher grade or more aggressive growth. The model of brain tumor growth was recently modeled to be a complex system of partial differential equations (PDEs). This system of PDE provides a significant amount of information about the grade of the tumor that cannot be easily measured from the tumor mass. For example, the PDEs have constants that are related to motility, angiogenesis and the mitotic rate; however, a single measurement of the brain tumor mass cannot easily be connected to these constants. This thesis presents a machine learning-based approach to estimate these parameters using a sequence of tumor mass measurements with an LSTM and a 1-D CNN network to solve the inverse problem of PDE parameter estimation. Experimental results show that through working with different architectures, for both of two models, accuracy increases as a function of the number of tumor measurements. This result shows that the LSTM and 1-D CNN provide better and better approximations to these PDE parameters as more measurements are made available (e.g., a tumor mass measurement every month). Moreover, the perplexity can be used for LSTMs to effectively detect a change in the tumor grade. These findings demonstrate the potential that machine learning and neural networks can provide the medical community for solving the inverse problem of PDE parameter estimation.
Degree ProgramGraduate College
Electrical & Computer Engineering