MAGnitude-Image-to-Complex K-space (MAGIC-K) Net: A Data Augmentation Network for Image Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning excels in MRI reconstruction but requires extensive multi-coil k-space data, often unavailable in routine scans.
• Existing magnitude images are abundant, presenting an opportunity for data augmentation in deep learning models.
• Current data augmentation methods lack the realism needed for effective deep learning-based MRI reconstruction.

Purpose of the Study:

• To introduce the MAGnitude Images to Complex K-space (MAGIC-K) Net for generating realistic multi-coil k-space data.
• To enable deep learning reconstruction using readily available magnitude images and limited raw k-space data.
• To improve the performance and generalizability of deep learning MRI reconstruction.

Main Methods:

• Developed the MAGIC-K Net, a deep learning model to synthesize multi-coil k-space data from anatomical DICOM magnitude images.
• Utilized pairs of anatomical images to generate realistic intensity variations and displacements.
• Compared MAGIC-K Net augmentation with basic data augmentation techniques.

Main Results:

• MAGIC-K Net generated more realistic data augmentation compared to basic methods.
• High-resolution Diffusion Weighted Image (DWI) reconstruction showed significant benefits from MAGIC-K Net augmentation.
• Deep learning reconstruction performance was enhanced qualitatively and quantitatively in both healthy volunteers and tumor patients.

Conclusions:

• The MAGIC-K Net effectively addresses the data scarcity issue for deep learning MRI reconstruction.
• This method facilitates superior image reconstruction, particularly for DWI, by leveraging existing magnitude images.
• MAGIC-K Net improves the clinical applicability of deep learning for diverse patient populations and pathologies.