Recent advances in machine learning, specifically deep learning, are poised to transform imaging focused healthcare domains including radiology and ophthalmology. In this seminar, I will discuss our recent work that spans the spectrum from algorithm development to clinical applications to deployment in the clinic, and the challenges and opportunities in each phase. Following a brief review of exemplar applications built using multi-institutional, multi-national datasets, I will highlight many of the practical challenges of applying AI on “real world” datasets that have motivated our technical innovations.

A discussion of some of the reported limitations of deep learning, including brittleness, catastrophic forgetting, and  the potential for biases, will be followed by mitigation strategies.  Variability and bias, even among experts, can lead to noisy labels available for training models, especially for diseases that lie on a spectrum. Using data from our user studies, I will briefly discuss technical approaches to creating severity scales based on pairwise comparisons and Siamese networks and highlight their applications to a number of clinical problems including the analysis of fundus photographs in retinopathy of prematurity and chest X-rays in COVID-19. I will also highlight the potential that AI has in health disparities research and monitoring of health systems.  Motivated by challenges in sharing healthcare data, I will discuss some of our recent work in this area of collaborative learning including federated learning, cyclical weight transfer and split learning.

Finally, I will conclude with our vision for our future research in this nexus between clinical needs and technical advancements.