AutoSamp: Autoencoding k-Space Sampling via Variational Information Maximization for 3D MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated MRI protocols often use fixed k-space undersampling patterns, limiting reconstruction quality.
• Optimizing these sampling patterns is crucial for improving MRI image fidelity but remains a complex challenge.

Purpose of the Study:

• To develop a novel deep learning framework, AutoSamp, for the joint optimization of MRI k-space sampling patterns and image reconstruction.
• To enhance the quality and sharpness of accelerated MRI scans through data-driven sampling optimization.

Main Methods:

• AutoSamp utilizes variational information maximization for joint optimization.
• The framework employs a non-uniform Fast Fourier Transform encoder for continuous k-space sample location optimization on a non-Cartesian plane.
• A deep reconstruction network serves as the decoder.

Main Results:

• AutoSamp demonstrated superior reconstruction quality compared to variable density and Poisson disc sampling methods across various acceleration factors (R=5, 10, 15, 25), achieving significant PSNR improvements.
• Prospectively acquired accelerated 3D FSE sequences using AutoSamp's optimized patterns showed enhanced image quality and sharpness.
• Learned sampling patterns adapt to factors like acceleration, noise, anatomy, and coil sensitivities, influencing sampling density and k-space coverage.

Conclusions:

• AutoSamp offers an effective deep learning approach for optimizing MRI sampling patterns and reconstruction simultaneously.
• The data-driven optimization framework leads to improved image quality in accelerated MRI acquisitions.
• The study highlights the influence of various factors on the learned sampling patterns, providing insights into adaptive MRI acquisition.