UCLA

Medical imaging is a class of imaging technology to understand the human's body by non-invasively creating visual representations. Despite its ubiquitous use, diagnosis from medical imaging remains an uncertain and time-costly process for physicians. As such, there exists a call for automatic tools that can provide assistance to physicians in analyzing medical imaging data. Although current AI solutions can achieve promising diagnostic accuracy with deep learning in many applications, there has been a resistance to adopt AI-based diagnosis in clinics. We believe this is primarily due to the lack of designs that center the tools on the needs of physicians. In this dissertation, we explore user-centered deep learning for medical image analysis. In specific, we are interested in two research questions. First, how to use the concept of user-centered design to drive the development of deep learning (DL) algorithms? In specific, contrary to most existing computer aided diagnosis (CADx) systems, what are the ways of applying DL in clinical settings besides simply providing the predicted diagnostic results? This is because AI-based diagnosis can still raise ethics and safety concerns in the foreseeable future by considering the imperfectness of AI and the high stake of medical decision making. To answer this question, we perform formative studies to understand physician's needs, based on which we formulate novel deep learning tasks and provide pioneer solutions. Our works covers a wide range of medical domains of neural-radiology, dentistry, and forensics. Second, how to design the interactions between tools and their users so that they can be seamlessly integrated into users' workflow? To answer this question, we build interactive CADx systems with deep learning algorithms embedded, and perform comprehensive user studies to understand the designs. We experiment with a visualization tool for dentists to perform pre-surgical patient education, and a dental health monitoring tool for layman users. We conclude the dissertation by discussing the current major challenges for user-centered AI tools that we learnt from our studies.