Predicting Future Tumor Location in Patients with Brain Metastases; Criteria for Successful Integration of Machine Learning Tools in a Medical Setting

The technical portion of this thesis aims to predict tumor formation prior to visualization using machine learning technologies. In order to complete this research, T2 contrast MRI data was collected from the University of Virginia Gamma Knife Radiosurgery Lab. These data sets included patients with multiple brain metastases, as well as others with no brain metastases. Due to the aggressiveness of the cancers GKS is intended to treat, for many patients imaging is available prior visualization of any specific metastasis. Metastasis volume was extracted from imaging prior to tumor visualization and compiled into a precancerous tissue data set. In contrast, this same approximate volume was extracted from the same relative location in a healthy brain and compiled into a healthy tissue data set. Both data sets were inputted into the existing neural network AlexNet to train the algorithm to detect healthy and precancerous tissue. The resulting system when presented with a new MRI image differentiates between healthy and precancerous tissue. Precancerous tissue is then flagged for future monitoring by oncologists. By flagging these areas, the hope is that metastasis can be detected prior to significant growth, and therefore, the amount of radiation necessary to treat the cancer and the risks associated with this treatment can be decreased significantly.
Machine learning technologies, such as that described above, has shown promise in improving patient care. However, to realize this benefit at a larger scale, machine learning technologies must be successfully implemented in the field. Four major questions are addressed to determine the steps necessary for successful implementation: Who is affected by the technology? What factors effect individual acceptance of the technology over the current approach? What steps can be taken to ease integration of the technology? What are unintentional consequences of data collection by the technology? These questions are evaluated in the context of research organized by policy analysis, historical case studies, interviews with medical professionals, and discourse analysis. To guide research and asses the relationship between machine learning technologies and society, Actor-Network Theory, technological momentum, and the concept of a paradigm shift are utilized as STS frameworks. This paper defines specific criteria for successful implementation of machine learning technologies in a medical setting. Defining these criteria provides a structed system that allows for society to benefit from the utility these machine learning technologies provide. Additionally, the knowledge acquired while defining these criteria establishes a relationship between society and an emerging technology that can be utilized to inform future STS analysis.

School of Engineering and Applied Science
Bachelor of Science in Biomedical Engineering
Technical Advisor: William Watkins
STS Advisor: Bryn Seabrook
Technical Team Members: Pamela Flintsch Medina, Connor Grubbs