Deep, deep learning with BART.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Magnetic Resonance Imaging (MRI) is crucial for medical diagnostics.
• Reproducible research in MRI image reconstruction is essential for clinical translation.
• Existing reconstruction methods often lack flexibility for advanced deep learning integration.

Purpose of the Study:

• To develop a deep-learning-based image reconstruction framework for reproducible research in MRI.
• To extend the BART toolbox with automatic differentiation capabilities for gradient computation.
• To facilitate the implementation and training of advanced deep learning reconstruction networks.

Main Methods:

• Extended the BART toolbox with a nonlinear operator framework enabling automatic differentiation.
• Integrated existing MRI-specific operators (e.g., nonuniform fast Fourier transform) and neural network building blocks.
• Implemented two state-of-the-art unrolled reconstruction networks, Variational Network and MoDL, within the framework.

Main Results:

• The BART-based framework allows construction and training of deep image-reconstruction networks using gradient-based optimization.
• Achieved comparable performance in training time and reconstruction quality to original TensorFlow implementations.
• Demonstrated the framework's capability for advanced deep learning-based reconstruction.

Conclusions:

• A general framework for deep-learning-based reconstruction in MRI has been established by integrating nonlinear operators and neural networks into BART.
• This framework supports reproducible research by providing a unified platform for developing and testing reconstruction algorithms.
• The developed framework offers a powerful tool for advancing MRI image reconstruction.