J-MoDL: Joint Model-Based Deep Learning for Optimized Sampling and Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Modern Magnetic Resonance Imaging (MRI) utilizes compressed sensing and deep learning to shorten scan times by reconstructing data from undersampled measurements.
• Image quality in these accelerated MRI techniques is critically dependent on the chosen sampling pattern.

Purpose of the Study:

• To introduce a novel continuous strategy for jointly optimizing MRI sampling patterns and deep learning network parameters.
• To enhance the image quality and efficiency of accelerated MRI acquisition.

Main Methods:

• Developed a multichannel forward model incorporating a non-uniform Fourier transform with continuously defined sampling locations.
• Integrated this model into the data consistency block of a model-based deep learning image reconstruction framework.
• Facilitated joint and continuous optimization of both sampling patterns and Convolutional Neural Network (CNN) parameters.

Main Results:

• Demonstrated that joint optimization of sampling patterns and reconstruction modules significantly improves the performance of deep learning-based MRI reconstruction algorithms.
• Achieved superior image quality compared to methods with fixed or independently optimized sampling patterns.

Conclusions:

• The proposed joint optimization strategy offers a powerful approach to enhance accelerated MRI.
• This method holds potential for improving diagnostic accuracy and patient comfort in clinical MRI settings.