Imaging and Biomedicine AI

Hui Xue head shot

Senior Investigator Research Interests

Research Interests
  • Artificial intelligence technology and system to advance multi-modality clinical and biomedical imaging.
  • Understand the mechanism for how the deep learning models work in imaging and study the algorithms to maximize the nonlinear performance boost.
  • Novel deep learning models and training methods to build the imaging foundation model and apply them to advance multiple biomedical imaging modalities.
  • Imaging reconstruction methods using the deep learning.
  • Novel analysis method to enable zero or few-shot automated quantification of large clinical and biomedical datasets, using foundation model and self-supervised learning.
  • Build and deploy deep learning systems to assemble data flywheels.

Example projects

AI global network

global map with points on different hospitals that use the AI applications they developed.

AI applications developed by me had been globally deployed to 70 hospitals (May 2022) and used daily on clinical patients (~50K patients annual). To achieve this scale of deployed AI, a software stack was developed to support all steps of applied AI development, from product definition, data collection, model training and management, testing and deployment to end-user.

Deep Learning based Cardiac Perfusion Quantification – “AI Doctor”


An image of a heart scan with text below saying "An AI doctor is analysing heart scans in dozens of hospitals"Cardiac disease is a leading cause of death worldwide. I developed a novel AI application to direct measure blood supply in the myocardium and reported analysis and disease prediction on the MR scanners. This AI application is fully automated, thanks to robustness provided by deep learning and big data. This technique is deployed worldwide through our AI global network and used daily on patients. 

This application was featured in many media news. On the left is the news reported on NewScientist, “An AI doctor is analysing heart scans in dozens of hospitals.”

Read more:



Distributed MR Imaging AI on Azure Cloud

Image of cloud computing process. From left to right, Scanner, then Raw signal and images measurements reports. THen SSH jump server. Then move to vnet where there is an image of a cloud with VM1 and VM2 inside. Lastly move over to box icon with text reading "Storage account shared by all pods."

Cloud computing is the key to clinically deploy advanced imaging computing and AI models to enable fast and robust MR imaging. Based on Gadgetron, a motion correction, non-linear compress sensing reconstruction was developed to run on the Azure cloud. By parallelizing the computation over multiple nodes, 10x speedup is achieved to reconstruct the entire dataset to cover the heart in 1min. After the reconstruction, the automated image analysis was achieved by deploying AI models to segment and quantify the cardiac contractibility. This application is deployed to two data centers, serving 10 hospitals in both Europe and North America and running every day to help clinicians and patients. 

Microscopy Imaging AI using Transformer

Image of Microscopy Imaging AI using Transformer process. Begins with Transformer Backbone, then moves to Fine-tuning, then to A modiel adapted to new experiment then lastly Inference.

Due to the large variation in microscopy types, vendors, experimental setups and cell/organelles, microscopy AI typically requires time-consuming training separate models for every experiment. To improve generalization, enable fast training adaptation and to reuse accumulated data, the conventional transformer architecture was modified to efficiently process large-format microscopy images. A backbone model built with this modified transformer module was developed to enhance the microscopy images and show strong ability to be adapted to new experiments.

Deep Super Imaging

Image of Deep imaging process. An image of a transparent cube with Physics listed at the top. In the middle the text in a box reading "Information from previous scans" with a right angle behind it.

Current clinical and biomedical imaging (e.g. MRI, CT, light microscopy etc.) is built on top of imaging physics and signal processing algorithms. One example is the MRI where pulse sequences (physics) and parallel imaging (signal processing) are used together to achieve required tissue contrast, resolution and SNR. While this scheme is well established, it only uses the data or information acquired in this current scan. With deep learning, it is possible to significantly boost imaging efficiency by  training models on large datasets previously acquired to help the current scan. This scheme adds a new axis and great potential to achieve superb imaging efficiency. The MRI LGE images on the right is reconstructed with deep learning models trained with ~200K prior scans. Learning based imaging is clearly superior to the current linear imaging technique on the left.

Meet the Team

Hui Xue head shot

Hui Xue, PhD

Director, Imaging and Biomedicine AI Program

Dr. Xue is an active researcher and developer in deep learning for vision, imaging and biomedicine applications. AI solutions he developed are deployed to more than 70 hospitals globally to help clinicians and patients. His research focuses on developing novel AI technique to enable new biomedical research. He is currently leading the Imaging AI program at the National Heart, Lung and Blood Institute, NIH, to develop cutting-edge full-stack solutions from multiple modalities

He earned his PhD from Imperial College London in 2008 and got the MSc in Biomedical Engineering and BSc in ECE from Xi’an JiaoTong University, China in 2003 and 2001. From 2008 to 2013, he worked as a research scientist in Siemens Corporate Research, Princeton, USA. He had authored over 100 scientific publications and holds 15 US patents.

For his contribution, he was co-awarded the Orloff Innovations Award in 2016 and 2020 by the Division of Intramural Research, NHLBI, for developing high performance image reconstruction software and for developing and deploying AI cardiac perfusion quantification. He is a Fellow of the Society for Cardiovascular Magnetic Resonance (FSCMR) and serving as editors and regular reviewers for several research journals.