Loading...

Deep Learning for Compressed Sensing MRI Reconstruction

Saleminejad, Abbas | 2021

24 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54139 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Vosoughi Vahdat, Bijan; Fatemizadeh, Emad
  7. Abstract:
  8. Medical imaging is an indispensable component of modern medical research as well as clinical practice. However, Magnetic resonance Imaging and Computational tomography are expensive, and it is difficult to be used in many scenarios worldwide. As such, many parts of the world still do not have sufficient accessibility to these techniques. To make medical devices more accessible, affordable and efficient, it is crucial to reflect upon our current imaging paradigm for smarter imaging. According to Compressed Sensing theory, if there is information other than the signal bandwidth value, such as the being sparse in a suitable domain, nonlinear optimization methods can be used to accurately reconstruct the signal from observations far less than the Nyquist theory states. This theory can be used to reduce the scan time on MRI, which is important in medical imaging.In this dissertation, the theory of compressed sensing was briefly proposed and, therefore, a deep generative adversarial network was used to reconstruct MR brain images. In this work, convolutional networks have been used to implement different parts of the network. Finally, the generator receives images with noise and artifacts and return the denoised output. According to the PSNR evaluation criteria, compared to classical networks such as TV and ADMM, as well as deep learning methods such as WGAN and DCCNN, this network had improvements in PSNR criteria. The results show that our proposed network has been able to have 2% improvement in comparison to deep learning networks and a 9% improvement over the TV network. Another advantage of the proposed networks over mathematical and classical works that use iterative methods is the very short reconstruction time, the average reconstruction time of our method is 30 ms. SARA-GAN network could able to improve the quality of images among current networks, But there are drawbacks to this network such as the large number of training parameters, high training and reconstruction time, stability as well as the need for more data and advanced equipment for training which limits the usage of this network
  9. Keywords:
  10. Nonlinear Optimization ; Generative Adversarial Networks ; Magnetic Resonance Imagin (MRI) ; Compressive Sensing ; Sparse Domain Transformation ; Image Reconstruction ; Medical Imaging

 Digital Object List