Cardiac CT is the first line imaging modality for diagnosis of cardiovascular diseases. A major challenge of cardiac CT remains motion artifacts due to fast and/or irregular cardiac dynamics. The existing motion artifact suppression algorithms can be improved based on distribution shifts due to anatomical and pathological variations in patients, protocol and technical changes of scanners, and other factors. In this paper, we construct a diversified dataset consisting of over 1,000 cardiac CT images of diverse features. Also, we provide a pipeline for source-agnostic vessel segmentation and motion artifact scoring. Our results demonstrate the merits of the approach and suggest a guideline for ensuring source-agnostic representativeness of anatomical and pathological imaging biomarkers in cardiac CT applications and beyond.
|