eduzhai > Applied Sciences > Engineering >

Reduce slice spacing of MR images by super-resolution learned without ground-truth

  • KanKan
  • (0) Download
  • 20210430
  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: High-quality magnetic resonance (MR) image, i.e., with near isotropic voxelspacing, is desirable in various scenarios of medical image analysis. However,many MR acquisitions use large inter-slice spacing in clinical practice. Inthis work, we propose a novel deep-learning-based super-resolution algorithm togenerate high-resolution (HR) MR images with small slice spacing fromlow-resolution (LR) inputs of large slice spacing. Notice that most existingdeep-learning-based methods need paired LR and HR images to supervise thetraining, but in clinical scenarios, usually no HR images will be acquired.Therefore, our unique goal herein is to design and train the super-resolutionnetwork with no real HR ground-truth. Specifically, two training stages areused in our method. First, HR images of reduced slice spacing are synthesizedfrom real LR images using variational auto-encoder (VAE). Although thesesynthesized HR images are as realistic as possible, they may still suffer fromunexpected morphing induced by VAE, implying that the synthesized HR imagescannot be paired with the real LR images in terms of anatomical structuredetails. In the second stage, we degrade the synthesized HR images to generatecorresponding LR images and train a super-resolution network based on thesesynthesized HR and degraded LR pairs. The underlying mechanism is that such asuper-resolution network is less vulnerable to anatomical variability.Experiments on knee MR images successfully demonstrate the effectiveness of ourproposed solution to reduce the slice spacing for better rendering.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...