(This is the research page for Jingyan Xu, Ph.D., Associate Professor of Radiological Physics Division, Department of Radiology.)

X-ray CT image formation refers to the data processing chain, that starts from a patient entering the CT scanner, going through the data acquisition and image reconstruction process, and ends with the presentation and interpretation of a CT image that would provide the answer both doctors and the patient are looking for, e.g., the cause of a discomfort.

Our research involves developing and improving different components in the CT image formation process, including data acquisition, image reconstruction, and image interpretation to improve diagnosis and patient care. Some recent and on-going projects are listed below.

Spectral data acquisition on energy-integrating CT using a Stationary Spatial Spectral Encoder (S3E)

We introduce the stationary spatial spectral encoder (S3E), a simple hardware addition to energy integrating CT systems to enable multi-energy data acquisition. The S3E is made of a ring of spectral filters that can be affixed to the CT gantry and provides continuous modulation of the x-ray source spectrum. The S3E can emulate many existing multi-energy CT solutions; moreover, the expanded surface area of spectral materials enables new spectral CT methods.  This work was recently presented at the 8th CT meeting 2024 (https://www.ct-meeting.org/) in Bamberg, Germany, where we discussed the design concept of the S3E, and used a few examples to illustrate its potential applications.

Characterizing the noise and resolution properties of DL CT image formation

Deep-learning (DL) based CT image generation methods are often evaluated using RMSE and SSIM. By contrast, conventional model-based image reconstruction (MBIR) methods are often evaluated using image properties such as resolution, noise, bias. Calculating such image properties requires time consuming Monte Carlo (MC) simulations. For MBIR, linearized analysis using first order Taylor expansion has been developed to characterize noise and resolution without MC simulations. This inspired us to investigate if linearization can be applied to DL networks to enable efficient characterization of resolution and noise. More details can be found in our recent publication.   https://doi.org/10.1109/TMI.2022.3214475  

Task-based measures of image quality (IQ) using ROC/AUC analysis 

DL CT image generation is undergoing rapid growth. Image quality (IQ) evaluation of these DL methods is almost exclusively based on RMSE, PSNR, and SSIM. These IQ metrics are easy to compute and work well in computer vision. Medical images, unlike natural images such as cars or trucks, are created to serve a specific purpose or task. In diagnostic CT, the task is most often the detection of an abnormality, e.g., a lesion.  Metrics such as PSNR and SSIM are known to be poor predictors of diagnostic task performance, casting doubt on the champion algorithms’ clinical utility. We work on developing and applying task-based measures of IQ for DL image formation and the associated ROC/AUC methodology.  The following is our recent work on multi-class ROC analysis. https://doi.org/10.1016/j.neucom.2024.127520

…