ACCELERATED PARALLEL MRI USING MEMORY EFFICIENT AND ROBUST MONOTONE OPERATOR LEARNING (MOL).

Area of Science:

• Medical Imaging
• Machine Learning
• Applied Mathematics

Background:

• Model-based deep learning enhances parallel MRI acceleration by integrating imaging physics with learned priors.
• Existing methods like unrolled algorithms face memory efficiency challenges.

Purpose of the Study:

• To evaluate the effectiveness of the monotone operator learning (MOL) framework for parallel MRI acceleration.
• To compare MOL's performance against established unrolled algorithms in MRI.

Main Methods:

• The MOL algorithm utilizes a gradient descent step with a monotone convolutional neural network (CNN).
• It incorporates a conjugate gradient algorithm to ensure data consistency.
• The approach alternates between these steps for image reconstruction.

Main Results:

• MOL demonstrates comparable guarantees to compressive sensing, including uniqueness, convergence, and stability.
• The method proves significantly more memory-efficient than traditional unrolled deep learning techniques.
• Validation in static and dynamic parallel MRI settings shows competitive performance.

Conclusions:

• The monotone operator learning framework is a viable and efficient tool for accelerated parallel MRI.
• MOL offers a promising alternative to existing methods, balancing performance with reduced memory footprint.
• This approach has potential for improving MRI scan times and image quality.