Compressed Sensing: From Research to Clinical Practice with Deep Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Compressed sensing (CS) enables high-resolution image reconstruction from undersampled MRI data, potentially reducing scan times and improving patient experience.
• However, traditional CS methods face challenges including artifacts, extensive parameter tuning, and long reconstruction times, hindering clinical adoption.

Purpose of the Study:

• To review classical compressed sensing (CS) formulation.
• To outline the transformation of CS into a deep learning-based reconstruction framework using unrolled neural networks.
• To discuss clinical application considerations for these advanced MRI reconstruction techniques.

Main Methods:

• Review of classical compressed sensing (CS) principles.
• Introduction to unrolled neural networks for learning complex image priors.
• Demonstration using open-source Python code and open databases.

Main Results:

• Unrolled neural networks offer a practical approach to address CS limitations in MRI reconstruction.
• Deep learning integration allows for learning complex image priors from historical data.
• The tutorial provides a framework for transforming traditional CS into deep learning-based methods.

Conclusions:

• Unrolled neural networks represent a significant advancement in overcoming challenges associated with compressed sensing MRI.
• This deep learning approach facilitates faster, more accurate, and clinically viable MRI reconstructions.
• The tutorial equips researchers and clinicians with knowledge and tools for applying these methods in practice.