Motion compensation in cardiac CT for the entire heart using a residual U-Net

Cardiac motion remains a major challenge in maintaining image quality in cardiac CT. Despite continuous progress in temporal resolution and scan protocol optimization, there are still cases where the heart moves up to several millimeters during a scan. This can result in severe motion artifacts in the corresponding CT reconstructions, diminishing the diagnostic value of the scan. A prevalent method to address these artifacts and enhance temporal resolution is based on so-called partial angle reconstructions (PARs). In this approach, the scan is divided into multiple angular intervals, each reconstructed individually. Motion between these PARs is then corrected by applying motion vector fields (MVFs). After deforming all PARs to the same reference motion state using these MVFs, they are summed to produce the final, artifact-free reconstruction. Unlike existing techniques that focus on small areas around the coronary arteries, this work expands the strategy to encompass the entire heart. A residual U-net is trained to predict MVFs based for a given PARs. When applied to test data, the proposed method effectively eliminates nearly all motion artifacts across various heart regions, including the coronary arteries and cardiac valves. Consequently, the average mean absolute error (MAE) for the entire heart, in comparison to a motion artifact ground truth, was reduced from 32.74 HU to 8.37 HU.