Global and local feature extraction based on convolutional neural network residual learning for MR image denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Magnetic Resonance (MR) images exhibit varying noise distributions globally and locally.
• Existing convolutional neural networks struggle to simultaneously preserve global structure and local details in MR image denoising.

Purpose of the Study:

• To develop an advanced convolutional neural network for denoising 3D MR images.
• To enhance the preservation of both global structure and local details in denoised MR images.

Main Methods:

• Proposed a parallel and serial network (3D-PSNet) for 3D MR image denoising.
• Utilized residual depthwise separable convolution for local feature learning and parameter efficiency.
• Employed residual dilated convolution for global feature extraction and expanded receptive fields.
• Integrated reinforced residual convolution blocks with dense connections for feature refinement.

Main Results:

• 3D-PSNet achieved high performance metrics: 47.79% peak signal-to-noise ratio, 99.81% structural similarity index measure, and 0.40% root mean square error.
• Demonstrated competitive denoising effects across three public datasets.
• Ablation experiments validated the effectiveness of all proposed modules.

Conclusions:

• 3D-PSNet effectively restores global structure and local fine features in MR images by leveraging multi-scale receptive fields and residual dense connections.
• The proposed method is expected to aid physicians in rapid and accurate patient diagnosis.