POCS-Augmented CycleGAN for MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Magnetic Resonance (MR) image reconstruction is computationally intensive.
• Deep learning (DL) shows promise for accelerating MR image reconstruction.
• Traditional methods and DL approaches are often considered separate.

Purpose of the Study:

• To investigate the complementary nature of traditional MR image reconstruction methods and DL.
• To develop a hybrid approach combining DL and traditional algorithms for improved MR image reconstruction.
• To evaluate the effectiveness of the proposed hybrid method against existing state-of-the-art techniques.

Main Methods:

• A hybrid DL method integrating a generative adversarial network with cycle loss (CycleGAN) and Projection Onto Convex Set (POCS) was developed.
• POCS updated the output of initial CycleGAN training iterations, providing enhanced data for subsequent training.
• The hybrid method was tested using sub-sampled Magnetic Resonance Imaging data.

Main Results:

• The hybrid CycleGAN-POCS method demonstrated superior MR image reconstruction quality.
• Performance was compared against other DL methods (U-Net, GAN, RefineGAN) and compressed sensing.
• The proposed approach yielded the best reconstruction results among all tested methods.

Conclusions:

• Traditional methods and DL are complementary for MR image reconstruction.
• The hybrid CycleGAN-POCS approach offers a promising direction for high-quality, accelerated MR imaging.
• This work validates the synergy between iterative reconstruction and deep learning for medical imaging applications.