Deep learning-based reconstruction (DRAC) for multi-phase cardiac CT angiography

The National Heart, Lung, and Blood Institute awarded the University of Texas at Arlington $579,745 on July 25, 2026, under the Cardiovascular Diseases Research program (CFDA 93.837) to develop deep learning-based reconstruction (DRAC) methods for multi-phase cardiac CT angiography. The project runs through June 30, 2029, with performance at the university's Arlington, Texas campus.

The award funds development of four-dimensional deep learning reconstruction algorithms that combine model-based iterative reconstruction with features learned from large high-quality multi-phase CT angiography image databases. The research targets ultra-low radiation doses for multi-phase cardiac imaging—approximately 3 millisievert or less using low tube current and fast-pitch scans—while maintaining or improving diagnostic accuracy for coronary artery disease assessment. The project includes construction of a synthetic patient library using computational human phantoms with real patient parameters to support algorithm training and validation.

This is a Project Grant, a competitive research award mechanism. The University of Texas at Arlington, a public research university and Hispanic-Serving Institution within the University of Texas System, serves as the recipient through its Office of Research Administration.

Abstract

Coronary computed tomography angiography (CTA) can noninvasively assess coronary artery disease (CAD), the number one killer in the world, and avoid risks from cardiac catheterization of invasive coronary angiographies. Although cardiac-gated multi-phase CTA (MP-CTA) provides more cardiac functional information and images with less motion artifact than single-phase CTA (SP-CTA), its high radiation dose is detrimental to patients, especially for 80% of them that are with negative findings. Reducing MP-CTA imaging dose to a level equivalent to, or less than SP-CTA will significantly increase the benefit-to-risk ratio of MP-CTA and to make a huge impact to CAD patient management. Lowering radiation dose in medical imaging has been vigorously pursued due to imaging prevalence and potential cancer risk of ionizing radiation. Although notable progress has been made by developing advanced reconstruction methods, such as model based iterative reconstruction (MBIR) and 4D reconstruction (3D space + 1D time), the dose reduction potential is far from its limit. The recent breakthrough in deep learning provides an unprecedented opportunity to leapfrog the dose reduction performance through the development of 4D deep learning-based reconstruction (DRAC) methods that combines MBIR and features learned from a large high-quality MP-CTA image database. This development will shift the conventional "individual-data-dependent" reconstruction paradigm to a revolutionary "individual-and-big-data-dependent" reconstruction paradigm. We hypothesize that 4D DRAC can achieve ultra-low dose for MP-CTA (about 3 MSV or less) using low tube current and limited angle (FAST-PITCH) scan with the same or improved diagnostic accuracy. To test this hypothesis, the following specific aims are proposed: 1) to build a synthetic patient library using computational human phantoms with real patient parameters and textures based on retrospectively collected real patient MP-CTA data for method development and hypothesis testing; 2) to develop image-domain DRAC (IDRAC) based on 3D and 4D deep learning and 4D DRAC methods using both CT imaging physics and features extracted by IDRAC; and 3) to quest the dose reduction limit of 4D DRAC using model-observer studies on synthetic patient data and to validate the findings using human observer studies on real patient data. These novel 4D DRAC methods can be readily expanded to different CT systems and other dynamic CT applications. The success of this project will substantially lower the radiation dose of MP-CTA, which will have a direct impact on the standard care of stenosis detection of CAD and a broad impact to its prognostic value of monitoring high-risk plaque progression. The outcomes of this project, such as synthetic and real patient MP-CTA data libraries and 4D DRAC algorithms, will provide valuable tools for the community of cardiac CT imaging. This area project will also provide valuable research opportunities in applying AI techniques on medical imaging to our large population of undergraduate students at UT arlington, a hispanic-serving institute.