Global attention-enabled texture enhancement network for MR image reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Convolutional Neural Networks (CNNs) show promise in accelerating Magnetic Resonance Imaging (MRI).
• Further research is needed to optimize CNNs for learning frequency characteristics and reconstructing texture in multicontrast MR images.

Purpose of the Study:

• To propose a novel Global Attention-enabled Texture Enhancement Network (GATE-Net) for highly undersampled MR image reconstruction.
• To enhance the learning of frequency characteristics and reconstruction of texture details in multicontrast MR images.

Main Methods:

• Developed GATE-Net incorporating a Frequency-Dependent Feature Extraction Module (FDFEM) and a Convolution-based Global Attention Module (GAM).
• FDFEM extracts high-frequency features from multicontrast images to improve texture.
• GAM utilizes the entire image receptive field to leverage beneficial shared information and suppress irrelevant information.

Main Results:

• Ablation studies confirmed the effectiveness of FDFEM and GAM.
• GATE-Net demonstrated superior performance across various acceleration rates and datasets.
• Quantitative metrics including peak signal-to-noise ratio and structural similarity were improved.

Conclusions:

• The proposed GATE-Net effectively reconstructs multicontrast MR images.
• GATE-Net achieves superior performance compared to existing state-of-the-art methods.
• The network is versatile, applicable to diverse acceleration rates and datasets.