Over the last two decades, image denoising has achieved significant performance improvements. However, the underlying idea of losing meaningful image information during filtering does not suit medical images well, where every voxel is essential. For highly radio-sensitive organs, such as hearts, modalities such as CT must use low-dose radiation. As a result, low-quality CT images are produced, making the diagnosis difficult. In myocardial perfusion imaging, cardiac imaging's residual motion also severely affects the diagnosis, requiring an accurate deformable registration algorithm. We propose a deep learning model to concurrently reduce noise and residual activities for low-dose myocardial CT perfusion without compromising the image information.
Article ID: 2021G05
Publisher: Canadian Artificial Intelligence Association