Research Projects

Machine learning has shown great potential in the medical image domain, but there are some limitations. The only test of robustness is the machine's successful matching of expert predictions, which can make errors hard to predict or correct. Additionally, the expertise developed by human experts is often used inappropriately for labeling data, wasting cognitive resources and training. The current project aims to develop a machine learning framework that employs expert cognitive processes while retaining the efficacy of the deep learning paradigm. By using insights from recent breakthroughs in vision and biology, the team hypothesizes that by interviewing expert radiologists about their judgments and cognitive processes, they can determine a "neural code" or latent space for medical diagnostics of specific pathologies. This would demonstrate the combination of expert cognition and machine "cognition," reduce the need for large resources for training and prediction, and create a more compact and efficient machine learning system with smaller training sets. This would be particularly important in rare medical conditions. Finally, this would allow for a full-fledged application and framework where it is possible to query the machine learning system and obtain expert-understandable results, an ability missing in current Artificial Intelligence research. A training tool for machine diagnostics could guide trainees on expert processes and not just expert results.