Interactive MRI Diagnostics: Improving CNN Predictions using Doctor Feedback; The Competition for the Future of Automated Medical Diagnostics

Gupta, Ankit, School of Engineering and Applied Science, University of Virginia
Basit, Nada, EN-Comp Science Dept, University of Virginia
Francisco, Pedro Augusto, EN-Engineering and Society, University of Virginia

My capstone research addresses the problem of doctors not being able to revise the brain tumor segmentation predictions made by machine learning models. Although several automated systems have the ability to take in some Magnetic Resonance Imaging (MRI) input, then predict whether a given user has a brain tumor, these systems don’t allow doctors to contribute their knowledge for cases when the automated system does not predict correctly. The technology that I created to address this problem was a website where doctors can visualize the predictions made by the machine learning model, then draw on the MRI image to revise the segmentation prediction to be more accurate. This website also allows the MRI segmentation image to be re-labeled, which can improve the machine learning model’s accuracy. It is important to consider the human and social dimensions of this technology, since this machine learning model is used by doctors in order to diagnose patients for brain tumor segmentation, which means that the perspectives of multiple groups of people are important. The main theories of Science, Technology, and Society (STS) that would apply to analyzing this approach are Actor-Network Theory and Interpretive Flexibility. Actor-Network Theory applies, because patients and doctors are connected to each other since doctors use the algorithm in order to make diagnostic decisions for patients. Interpretive Flexibility applies here, since different groups of people have different ways of interpreting the use of automated medical diagnostics tools. For example, doctors might interpret the use of these automated medical diagnostics tools to just be for simple diseases, while researchers might interpret the use to be for all types of diseases. The main method that I’m using to conduct my STS research is literature review, because this method allows me to look at research studies that cover the perspectives of a wide variety of stakeholder groups. These groups of stakeholders include patients, doctors, researchers, and insurers. The goal of my STS research is to better understand how automated medical diagnostics influences the perspectives of these different stakeholders. By comparing these perspectives, I expect to learn about key areas that developers should focus on improving. There are two main implications of my research. First, the main implication of my technical capstone research is that it allows doctors to contribute their expertise to machine learning models, while improving the accuracy of those models. In addition, the main implication of my STS research is that developers can better understand which problems to focus on solving in order to make sure that more people use automated medical diagnostics tools.

BS (Bachelor of Science)
machine learning, healthcare, automated medical diagnostics, interpretive flexibility, actor-network theory, magnetic resonance imaging

School of Engineering and Applied Science
Bachelor of Science in Computer Science
Technical Advisor: Nada Basit
STS Advisor: Pedro A. P. Francisco

All rights reserved (no additional license for public reuse)
Issued Date: