Accelerating 3D radial MPnRAGE using a self-supervised deep factor model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• 4D non-Cartesian MRI requires advanced reconstruction methods for high resolution and large parametric dimensions.
• Existing methods often struggle with computational demands and image quality in accelerated settings.

Purpose of the Study:

• To develop a self-supervised, memory-efficient deep learning method for 4D non-Cartesian MRI reconstruction.
• To improve image quality and quantitative accuracy in high-resolution MRI.

Main Methods:

• Developed the deep factor model (DFM) using neural networks and single-shot learning (SSL) from k-space data.
• Implemented a transfer learning (TL) approach for reduced reconstruction time.
• Compared DFM against subspace methods using phantom and in vivo MPnRAGE data for T1 imaging.

Main Results:

• DFM-SSL significantly improved image quality, reducing bias and variance in quantitative T1 estimates.
• DFM-TL decreased reconstruction time while maintaining performance comparable to DFM-SSL.
• Both DFM variants outperformed subspace methods in phantom and in vivo studies.

Conclusions:

• The DFM provides superior multicontrast image representation compared to subspace models, particularly in accelerated MPnRAGE.
• Self-supervised training is highly suitable for high-resolution, large-dimension MRI where deep learning training is computationally intensive.