JOINT OPTIMIZATION OF SAMPLING PATTERN AND PRIORS IN MODEL BASED DEEP LEARNING.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning (DL) offers advanced solutions for compressed sensing Magnetic Resonance Imaging (CS-MRI) by recovering images from undersampled data.
• Current DL methods lack theoretical understanding of image redundancies, hindering optimal sampling pattern selection for effective reconstruction.
• This gap complicates the choice of sampling strategies, impacting the quality of recovered MR images.

Purpose of the Study:

• To optimize sampling patterns and reconstruction parameters within a model-based DL framework for CS-MRI.
• To enhance the performance of image recovery from highly undersampled MRI data.
• To address the limitations of current DL approaches in understanding image properties and sampling effects.

Main Methods:

• Proposing a model-based deep learning framework for joint optimization of sampling patterns and reconstruction parameters.
• Implementing a strategy that decouples the effects of sampling and image properties during reconstruction.
• Comparing the proposed joint optimization approach against direct inversion Convolutional Neural Network (CNN) schemes.

Main Results:

• The model-based joint optimization strategy demonstrated superior performance compared to direct inversion CNN methods.
• Improved image recovery was achieved by effectively decoupling sampling and image properties.
• Quantitative and qualitative analyses confirmed the benefits of the proposed model-based scheme.

Conclusions:

• Joint optimization of sampling patterns and reconstruction parameters in a model-based DL framework significantly enhances CS-MRI performance.
• The proposed method offers a more theoretically grounded approach to CS-MRI, improving image quality and recovery.
• This work provides a pathway for developing more efficient and effective deep learning strategies for accelerated MRI acquisition.