Research
Examples of the many exciting projects in our group!
Diffusivity as a Biomarker for Knee Health
Osteoarthritis is the leading cause of disability and results in degenerated cartilage. It is known that contrast agents (from arthrography) diffuse into cartilage over time, with a rate dependent on cartilage health. Our hypothesis is that cartilage diffusivity can be quantified with PCCT and used as a biomarker for knee health. This leverages the high spatial resolution and accurate material quantification of PCCT, but still pushes the technical limits of PCCT resolution and quantification. For example, we are investigating the non-linear partial volume (NLPV) effect in photon counting detectors to further improve the spatial resolution.
Coronary Calcium Detection from Chest X-ray
Cardiovascular disease is the leading cause of death in the U.S. While CT is the gold standard for detecting coronary artery calcification (CAC), its cost and radiation limit screening. We show that single-shot dual-energy chest X-ray (DE-CXR) combined with an AI segmentation model enhances CAC visualization in simulations, pointing to a new path for early detection.
AI-Based CT Tube Current Modulation
Reducing radiation dose while preserving high image quality in regions of interest is a major challenge in CT imaging. Our Tube Current Modulation (TCM) optimization approach determines personalized TCM maps given scouts and scan range. Our prospective method was comparable to a fully retrospective approach. This approach allows patient-specific and task-driven imaging offering the best image quality at desired dose.
End-to-End Differentiable PCCT Imaging
Material decomposition (MD) is a critical path in quantitative PCCT imaging but depends on an iterative solver with pre-calibrated models (time-consuming and tedious) and relies on corrections that require calibration as well.
We propose a differentiable MD method:
1) Finding the gradients of the iterative MD solver, making the PCCT imaging chain differentiable.
2) Enabling data-driven cross-domain optimization and calibration to minimize human involvement (e.g., bin drift correction, scatter-correction neural network).
Printing Anthropomorphic CT Phantoms
Utilizing office laser printers to deposit high-attenuation toners on base materials enables the generation of low-cost customized 3D anthropomorphic CT phantoms that accurately replicate realistic anatomical structures and energy-dependent x-ray attenuation properties. We used an HP LaserJet 4100 printer with toner that contains iron oxide (i.e., high density and high effective atomic number), to print myocardial phantoms with infarcts of different contrast levels. The assembled phantom prototype was then scanned using a PCCT system.
Continuous Limited-Angle Dynamic 4DCT
To enable scanners with time-revolving ability, we developed a continuous limited-angle dynamic cone-beam CT reconstruction method with prior-driven unsupervised learning. This approach, called PriorDIP, combines the advantages of 4D prior-based iterative CT reconstruction with a deep image prior model without the need for an external training dataset. We demonstrated that PriorDIP faithfully presents the dynamics in a flow phantom study.
