Densely connected wavelet-based autoencoder for MR image reconstruction

Aghabiglou A., Ekşioğlu E. M.

45th International Conference on Telecommunications and Signal Processing, TSP 2022, Virtual, Online, Czech Republic, 13 - 15 July 2022, pp.212-215 identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/tsp55681.2022.9851354
  • City: Virtual, Online
  • Country: Czech Republic
  • Page Numbers: pp.212-215
  • Keywords: Deep learning, Densely Connected Residual Network, Magnetic resonance imaging, MR Image Reconstruction
  • Istanbul Technical University Affiliated: Yes

Abstract

© 2022 IEEE.Recently, methods based on deep learning have been introduced to the literature as a solution for accelerating magnetic resonance imaging technique. However, Image reconstruction from subsampled data is an ill-posed problem. In the current study, the wavelet package has been applied to deep networks. The replacement of the conventional downsampling and upsampling layers with Discrete Wavelet Transform (DWT) and Inverse Wavelet Transform (IWT) improved the reconstruction results. Moreover, the consequence of this substitution has been investigated on potent densely connected deep networks. The proposed novelty resulted in promising performance improvement in MR Image reconstruction.