Deep Adaptive Blending Network for 3D Magnetic Resonance Image Denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Magnetic Resonance Imaging (MRI) quality is vital for diagnosis and research.
• Image noise is a primary cause of quality degradation in MRI.
• Existing deep learning denoising methods struggle to integrate long-range, hierarchical, and 3D spatial information.

Purpose of the Study:

• To propose a novel deep learning model for enhanced MRI denoising.
• To address limitations in current methods by incorporating long-range, hierarchical, and 3D similarity information.
• To improve the visual quality and diagnostic utility of MRI scans.

Main Methods:

• Developed a Deep Adaptive Blending Network (DABN).
• Introduced a large receptive field residual dense block to capture long-range and hierarchical features.
• Implemented an adaptive blending method to utilize 3D MRI similarity for denoising, incorporating residual compensation.

Main Results:

• The proposed DABN effectively captures long-range dependencies and fuses hierarchical features.
• The adaptive blending method successfully leverages 3D MRI similarity for noise reduction.
• Experimental results demonstrate superior performance of DABN compared to state-of-the-art methods on clinical and simulated MRI data.

Conclusions:

• DABN offers a significant advancement in MRI denoising technology.
• The network's architecture successfully integrates multiple types of image information for superior noise reduction.
• The proposed method enhances MRI visual quality, benefiting clinical diagnosis and research.