Figure 1.

Overview of the proposed temporally aware volumetric GAN (TAV-GAN). The main component is a volumetric GAN (top). An ancillary temporal GAN (bottom), which is pre-trained, provides the temporally aware (TA) loss for the volumetric GAN training. Three objective functions, including content losses (SSIM, and L₁), adversarial loss, and TA loss, are used to train the volumetric GAN. The role of the content loss is to compel the volumetric generator to produce anatomically correct images, and the role of the TA loss is to compel the volumetric generator to produce temporally coherent image. The TA loss is calculated based on L₂ distance between features in two intermediate layers (Block 1 Conv 1 and Block 2 Conv 1) of the pre-trained temporal discriminator D^T when the output of the volumetric generator G^v and the ground truth image volumes are separately input to D^T. The temporal generator and discriminator take as input accelerated, aliased, and respiratory motion-corrupted magnitude 3D image patches from three consecutive temporal frames (t-1, t, and t+1), and produce an un-aliased, and respiratory motion-corrected 3D image patch for frame t.