Variable augmentation network for invertible MR coil compression.

Area of Science:

• Medical Imaging
• Deep Learning
• Data Compression

Background:

• Multi-coil Magnetic Resonance (MR) imaging generates large datasets, necessitating efficient compression techniques.
• Current coil compression methods often lack reversibility, limiting their application in medical scenarios where data integrity is crucial.

Purpose of the Study:

• To develop an efficient and reversible deep learning-based method for multi-coil MR data compression.
• To enhance the applicability of MR image compression in medical diagnostics through reversible recovery.

Main Methods:

• Introduction of the Variable Augmentation Network for Invertible Coil Compression (VAN-ICC), a deep learning algorithm.
• Utilizing normalizing flow-based models for inherent reversibility and applying variable augmentation to image/k-space variables.
• Training an invertible and bijective function for mapping data between original and compressed forms.

Main Results:

• VAN-ICC demonstrated effective and flexible compression on both fully-sampled and under-sampled MR data.
• Quantitative and qualitative comparisons showed VAN-ICC achieved significantly higher compression effects than traditional non-deep learning methods.
• The method successfully implemented reversible networks through variable augmentation technology.

Conclusions:

• VAN-ICC offers a competitive advantage over traditional coil compression algorithms, providing high compression efficiency with reversibility.
• The proposed deep learning approach enhances the potential for applying advanced compression techniques in medical imaging.