Memory efficient model based deep learning reconstructions for high spatial resolution 3D non-cartesian acquisitions.

Area of Science:

• Medical Imaging
• Machine Learning
• Magnetic Resonance Imaging

Background:

• Model-based deep learning (MBDL) faces GPU memory challenges for 3D non-Cartesian MRI reconstruction.
• High memory demand hinders the application of efficient techniques like gradient checkpointing for high-resolution 3D reconstructions.

Purpose of the Study:

• To develop a memory-efficient MBDL method for fast and high-quality 3D non-Cartesian MRI reconstruction.
• To overcome GPU memory limitations in MBDL for high-resolution 3D datasets.

Main Methods:

• Introduced block-wise learning, combining gradient checkpointing with patch-wise training.
• Decomposed volumes into patches, processed iteratively, and reconstructed the full volume for data consistency.
• Significantly reduced GPU memory requirements by linking memory usage to patch size instead of full volume.

Main Results:

• Achieved significant improvements in image quality compared to L1 wavelet compressed sensing.
• Reduced average reconstruction time by 38x.
• Successfully reconstructed highly undersampled 1.25 mm isotropic pulmonary MRA volumes on a single GPU.

Conclusions:

• Block-wise learning enables MBDL for high-resolution 3D non-Cartesian datasets.
• The method offers improved image quality and substantial reductions in reconstruction time.
• Facilitates advanced MBDL applications in challenging MRI scenarios.