CAREER: Deploying Transferable Medical Imaging Diagnosis System in Diverse Environments

Medical imaging, such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), chest X-ray, and retinal imaging, are valuable tools to assist in diagnosis. Medical imaging analysis has been significantly advanced using deep learning models. The knowledge extracted from large amounts of medical data can be used to make predictions for new patients. It has been demonstrated in many cases that the performances of machine learning models are comparable to board-certified radiologists or other professional experts, indicating the potential successful integration of those models in clinical applications. For example, imagine a patient notices a painless rash on their skin. If they could take a photo with a cell phone and receive a quick assessment comparable to experienced dermatologists, life-threatening diseases would be intervened or avoided early. However, the current success of deep learning is heavily dependent on large and high-quality labeled datasets. Such nearly perfect environments are only available in ideal lab environments because of the population shift, device differences, or rare diseases in real clinical applications. This project plans to focus on those specific challenges of non-ideal medical imaging diagnosis environments to advance the knowledge of building transferrable deep learning models and enhance national health by providing better tools for medical imaging diagnosis. Furthermore, this research will support the cross-disciplinary development of a cohort of Ph.D. and undergraduate students and outreach activities in the communities. Technically, this project will investigate and build transferable medical imaging diagnosis systems in diverse environments. The project proceeds with one overarching theme of leveraging the understudied geometric properties of deep neural networks to address three universal barriers when deploying medical imaging systems in various environments. Specifically, there are challenges to transferring models to novel classes, where there are not enough training samples, and novel domains, where the deploying environments change, and more importantly, preserving the previous knowledge in the model. If successful, the proposed research is expected to advance the understanding of building transferring deep learning models by leveraging a novel geometric interpretation of deep neural networks partitioning the input and feature space into generalized Voronoi diagrams. The driving applications of the proposed techniques are the prediction of long-tailed disease patterns on chest X-rays and ensuring consistent screening services for glaucoma in the community. In addition, the proposed methods have the potential to be extended to similar scenarios with a wide range of deployment environments. 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.