Pyramid Convolutional RNN for MRI Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Undersampled Magnetic Resonance Imaging (MRI) data requires fast and accurate reconstruction for clinical applications.
• Deep learning methods show promise but struggle with recovering fine details from undersampled data.

Purpose of the Study:

• To introduce a novel deep learning method, Pyramid Convolutional RNN (PC-RNN), for multi-scale MRI image reconstruction.
• To address the challenge of recovering fine details in undersampled MRI data.

Main Methods:

• Developed PC-RNN, a deep learning model utilizing three convolutional RNN (ConvRNN) modules for multi-scale feature learning.
• Implemented a pyramid-fashion combination of multi-scale reconstructions using a final CNN module.
• Handled multi-coil data directly as multi-channel inputs, avoiding coil sensitivity maps and using coil compression for standardization.

Main Results:

• PC-RNN demonstrated superior performance in recovering image details compared to other methods on fastMRI knee and brain datasets.
• The model achieved a coarse-to-fine image reconstruction through its multi-scale ConvRNN modules.
• The method was recognized as a winner in the 2019 fastMRI competition.

Conclusions:

• PC-RNN offers an effective approach for high-fidelity MRI reconstruction from undersampled data.
• The proposed method advances deep learning applications in medical imaging by improving detail recovery.
• PC-RNN provides a robust and efficient solution for clinical MRI practices.