DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Compressed sensing magnetic resonance imaging (CS-MRI) offers faster acquisition, reducing costs and patient discomfort.
• CS-MRI reconstructs images using less raw data, surpassing the Nyquist-Shannon sampling limit.
• Existing CS-MRI methods lack the integration of prior knowledge from large datasets.

Purpose of the Study:

• To develop a deep learning-based strategy for CS-MRI reconstruction.
• To bridge the gap between conventional single-image methods and large-dataset prior knowledge.
• To introduce a novel conditional Generative Adversarial Networks-based model (DAGAN) for CS-MRI.

Main Methods:

• Proposed a novel conditional Generative Adversarial Networks-based model (DAGAN) for CS-MRI reconstruction.
• Implemented a refinement learning method within a U-Net generator to stabilize reconstruction and reduce aliasing.
• Integrated adversarial and content loss, along with frequency-domain information, for enhanced texture and edge preservation.

Main Results:

• DAGAN demonstrated superior reconstruction quality compared to conventional and other deep learning CS-MRI methods.
• The proposed method effectively preserved perceptual image details, textures, and edges.
• Images were reconstructed rapidly, with each reconstruction taking approximately 5 ms, enabling real-time processing.

Conclusions:

• DAGAN offers a significant advancement in CS-MRI reconstruction, providing high-quality images with preserved details.
• The model's speed makes it suitable for real-time clinical applications.
• This deep learning approach effectively leverages prior knowledge for improved CS-MRI performance.