SCH: Generative Imaging Models for Verifying and Explaining Machine Learning Systems in Healthcare

Artificial intelligence - in particular, deep learning - is rapidly being developed for healthcare systems with great potential to improve disease diagnosis, treatment planning, and patient monitoring. However, the translation of these powerful models from research and development to everyday clinical use is being held back by the lack of trustworthiness in these systems. This project will explore strategies and develop methods for ensuring the robustness of deep-learning models in healthcare applications. A major obstacle to guaranteeing the behavior of deep-learning systems in healthcare is the wide variability in data across different healthcare sites, including a range of medical-imaging devices, data-collection protocols, and patient demographics. This can lead to data inputs to the system that are significantly different in nature from the data on which it was trained. To address this issue, we propose to develop robustness audits that assess how well a healthcare deep-learning system tailored to a specific site will operate at another site. Broader-impact aspects of the work include the potential to significantly and widely improve the effectiveness of deep learning in practical healthcare applications. Additionally, an array of educational and outreach activities are planned. The first goal of this project is to develop robustness audits using synthetic data that provide full coverage of test cases simulating conditions at a target healthcare site. This will be done by developing deep generative models with the ability to produce highly-realistic synthetic medical images that closely mimic the properties of imaging data collected at a target site. The second goal of this project is to use this generative-model framework to develop verification tools for measuring the robustness of a deep-learning healthcare system. This robustness will be expressed as regions in the latent space of the generative model, thereby restricting the set of data inputs to only valid medical images. The third goal of this project is to design natural-language models for communicating the results of the robustness audits to doctors. These models will produce textual descriptions of the conditions that would lead a deep-learning system to produce incorrect results at a particular site. Finally, the robustness-auditing system will be validated on real-world medical-imaging data in a cardiac-resynchronization therapy-planning task, where the deep-learning system to be audited predicts the optimal placement of pacemaker leads from magnetic-resonance imaging of the heart. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.