MRI reconstruction with enhanced self-similarity using graph convolutional network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel at fast Magnetic Resonance Imaging (MRI) reconstruction by utilizing local image information.
• However, CNNs may miss non-local image information due to limited receptive fields, impacting reconstruction quality.
• This study addresses this limitation by incorporating graph structures to capture long-range dependencies.

Purpose of the Study:

• To develop a novel network, Graph Convolutional Enhanced Self-Similarity (GCESS), for improved MRI reconstruction.
• To effectively integrate both local and non-local image information for more reliable image reconstruction.
• To enhance the structural integrity and detail preservation in reconstructed MRI images.

Main Methods:

• Reconstructing MRI images into a graph format to extract non-local self-similarity.
• Employing a hybrid approach combining spatial convolution and graph convolution within the GCESS network.
• Strengthening non-local similarities between image patches during the reconstruction process.

Main Results:

• The GCESS network demonstrated superior artifact suppression and detail preservation compared to state-of-the-art methods on in vivo knee and brain data.
• Quantitative analysis showed improved Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) for reconstructed images.
• Results were consistent across different sampling templates, validating the method's robustness.

Conclusions:

• The proposed GCESS network effectively combines spatial and graph convolutions for robust MRI image reconstruction.
• The method amplifies non-local self-similarities, significantly improving the structural integrity of reconstructed images.
• Experimental results confirm the superiority of GCESS in artifact reduction and detail preservation over existing methods.