AI-based Cardiac CT - Abstract Cardiovascular diseases (CVDs) are the leading cause of morbidity and mortality, with 19 million deaths globally in 2020 alone. Central to management of CVDs is deepening our knowledge of physiology and pathology of the heart. Major barriers to our greater understanding of the heart include its deep location and fast dynamics. Evidence from invasive coronary angiography indicates that the maximum velocity of cardiac structures is 52.5 mm/s, requiring a scan time of 19.1ms to eliminate motion artifacts. To achieve this temporal resolution with CT is extremely challenging. There have been substantial gains in CT hardware and software over the last decades (whole heart, dual-source, dedicated cardiac CT, and various approaches to cardiac motion compensation with ECG-gating) that have transformed coronary CT angiography into a robust and viable clinical tool. However, owing to the 140ms temporal resolution of current whole heart CT scanners, diagnosis is still challenging in patients who have irregular and/or fast heart rates especially in cases of arrhythmias and tachycardia, which commonly occur in older adults, many of whom exhibit atrial fibrillation. Here we will apply deep learning to radically improve cardiac CT reconstruction by attaining significantly higher spatial resolution, lower radiation exposure, and better image quality on both modern and legacy CT hardware. To improve wide-area-detector cardiac CT performance, we will develop a limited-angle reconstruction algorithm in the Analytic, Compressive, Iterative, and Deep (ACID) reconstruction framework that integrates a deep network trained on large data, sparsity-promotion, analytic modeling, and iterative refinement. For the first time, two of the preeminent advances in signal processing, compressive sensing and deep learning, will be combined to extract full information from scan data and image priors to freeze the beating heart. The specific aims H3 are: (1) Hyper Dataset: Projection datasets in the Radon space and the corresponding ground-truth images without motion artifacts in the image space will be generated in simulation, experiments, and clinical studies; (2) Hybrid Algorithm: A deep learning network and CS-module will be developed, integrated, and accelerated within the ACID framework for limited-angle free-breathing cardiac CT reconstruction, which will be shared on an open-source platform; and (3) Holistic Evaluation: The performance of our reconstruction software will be characterized, the stability and generalizability will be investigated, and task-based clinical applications will be demonstrated, including quantification of stenosis severity, aorta dimensions, and motion artifacts within the clinical setting of individuals with atrial fibrillation, tachycardia, and irregular heart rates. Completion of this project will yield a free-breathing cardiac CT algorithm with the unprecedented temporal resolution of 60ms, a 230% improvement over the state-of-the-art, allowing cardiac CT without clinically-relevant motion artifacts. This represents a major step towards the integration of model-driven and data-driven methods for CT image reconstruction, with a lasting impact on not only CT but also other tomographic modalities.
