A Digital Second Eye: An Ethnographic Study of Making Medical Imaging AI Development

 As technologies of machine learning and computer vision advance, a growing number of enterprises are applying these technologies to medicine. Based on an ethnography at an AI startup in South Korea, which is largely composed of scientists and doctors, I examine the process in which a medical imaging AI is shaped through a selective simulation of human vision and continuous negotiation between doctors and scientists. The development of medical imaging AI diagnostic software involves two steps: training the model with large sets of paired data of medical images and diagnostic information and then visualizing the result into a human-readable format. I consider the medical imaging AI as an attempt to produce a “digital second eye” and show how a doctor’s diagnostic ability is inevitably reduced to visual interpretation of medical images, while other medical expertise such as anatomical knowledge is excluded from the AI training process. Following the tensions and compromises between scientists and doctors in making the software, I then point out that simulation of a human eye allows the software to encompass not only codes and algorithms but also heterogeneous assumptions and understandings on medical knowledge, practice, and expertise. While scientists understand medical diagnosis as pattern recognition, which they believe can be better achieved with computers, doctors make medical diagnoses based on clinically proven medical experiences. Grounded in an empirical study of an ongoing project, this paper shows how disputes between scientists and doctors are provoked, mediated, and settled in the making of medical imaging AI.