Convolutional Recurrent Neural Networks for Dynamic MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Machine Learning
• Signal Processing

Background:

• Accelerating dynamic magnetic resonance imaging (dMRI) presents an ill-posed inverse problem, crucial for both signal processing and machine learning.
• Exploiting temporal correlations in MR sequences is key to resolving aliasing artifacts in accelerated dMRI.
• Traditional methods rely on iterative optimization algorithms, while deep learning offers a promising alternative for inverse problems.

Purpose of the Study:

• To develop a novel convolutional recurrent neural network (CRNN) for reconstructing high-quality cardiac MR images from highly undersampled k-space data.
• To jointly leverage temporal sequence dependencies and the iterative nature of traditional optimization algorithms for enhanced reconstruction.
• To improve the efficiency and accuracy of dynamic MRI reconstruction.

Main Methods:

• Proposed a unique CRNN architecture that embeds traditional iterative algorithm structures.
• Utilized recurrent hidden connections to model the recurrence of iterative reconstruction stages.
• Employed bidirectional recurrent hidden connections across time sequences to learn spatio-temporal dependencies.

Main Results:

• The proposed CRNN effectively learns temporal dependence and iterative reconstruction processes with a minimal number of parameters.
• Achieved superior reconstruction accuracy compared to current MR reconstruction methods.
• Demonstrated significant improvements in reconstruction speed.

Conclusions:

• The novel CRNN architecture offers an effective solution for accelerated dynamic MRI reconstruction.
• The method successfully integrates deep learning with iterative optimization principles for improved performance.
• This approach holds potential for faster and more accurate cardiac MRI acquisition.