Fast and Stable Neonatal Brain MR Imaging Using Integrated Learned Subspace Model and Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Neonatal brain MRI requires fast and stable imaging techniques.
• Deep learning accelerates MRI but needs substantial training data, which is scarce for neonates.

Purpose of the Study:

• To develop a fast and stable neonatal brain MRI method.
• To integrate a neonate-specific subspace model with deep learning for improved image reconstruction.

Main Methods:

• Utilized a subspace model to capture neonatal brain image features.
• Integrated the learned subspace model with a deep network for reconstruction from sparse k-space data.
• Employed model-driven deep learning for enhanced image quality.

Main Results:

• Validated the method's effectiveness and robustness on the dHCP dataset and multi-center data.
• Demonstrated remarkably stable reconstruction performance under perturbations.
• Showcased improved image reconstruction when the learned subspace was combined with other deep neural networks.

Conclusions:

• Subspace-assisted deep learning enables fast and stable neonatal brain MRI with sparse sampling.
• The proposed method shows potential to enhance the practical application of MRI in neonatal imaging.