Personalized synthetic MR imaging with deep learning enhancements.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Personalized synthetic MRI (syn-MRI) synthesizes MR images at various parameters from a few acquired images.
• Classical methods struggle with noise and ill-posed inverse problems in syn-MRI.
• Deep learning (DL) offers potential for improved spatial regularization, especially with limited personalized data.

Purpose of the Study:

• To develop and validate a DL-based pipeline for enhanced personalized 3D syn-MRI.
• To address limitations of classical least-squares (LS) and maximum likelihood estimators (MLE) in high-noise scenarios.
• To provide a practical workflow for syn-MRI with minimal training data.

Main Methods:

• A Deep Image Prior (DIP) with a U-net denoising architecture was employed for DL enhancement.
• A workflow, DIPsyn-MRI, was developed, utilizing DIP to enhance training images before parametric map estimation (LS/MLE).
• The method was implemented in a publicly available Python package, DeepSynMRI.

Main Results:

• DIPsyn-MRI demonstrated feasibility and improved performance on 3D datasets (spin-echo, FLASH) across various noise levels.
• DL enhancements improved syn-MRI quality, even with magnetic field intensity nonuniformity.
• Performance gains were observed for all but the very lowest noise levels.

Conclusions:

• The study provides a practical pipeline and software for DL-enhanced personalized syn-MRI.
• DIPsyn-MRI effectively improves image synthesis in challenging, data-limited personalized scenarios.
• This approach facilitates more accurate and robust personalized MR image generation.