Automatic Cardiac Magnetic Resonance Respiratory Motions Assessment and Segmentation

Cardiac magnetic resonance imaging (CMR) is a powerful non-invasive tool for diagnosing a variety of cardiovascular diseases. However, the quality of CMR imaging is susceptible to respiratory motion artifacts. Recently, an extreme cardiac MRI analysis challenge was organized to assess the effects of respiratory motion on CMR imaging quality and develop a robust segmentation framework under different levels of respiratory motion. In this paper, we have presented two different deep learning frameworks for CMR imaging quality assessment and automatic segmentation. First, we have developed 3D-DenseNet to assess the image quality, followed by 3D-deep supervision UNet with the residual module using pseudo labelling for automatic segmentation task. Experiments on the Challenge dataset showed that 3D ResNet with deep supervision using Pseudo Labeling with nnUNet achieved significantly better performance (8.747 LV, 3.787 MYO, and 5.942 RV) HD95 score than 3D-UNet.