A Multifeature Extraction Method Using Deep Residual Network for MR Image Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic resonance (MR) imaging is crucial for diagnostics but often suffers from noise interference, reducing image resolution and diagnostic accuracy.
• Deep learning models, particularly deep residual networks, have shown promise in image processing tasks, including denoising.

Purpose of the Study:

• To propose a multifeature extraction denoising algorithm based on a deep residual network to improve MR image resolution and reduce noise.
• To enhance the network's multiscale perception and optimize deep learning processes for clearer image generation.

Main Methods:

• A deep residual network incorporating a feature extraction layer with three different convolution kernel sizes for multiscale perception.
• Integration of batch normalization and residual learning to accelerate and optimize the deep network.
• A joint loss function combining perceptual loss and mean square error for pixel-level and semantic feature learning.

Main Results:

• The proposed algorithm demonstrated superior denoising performance compared to other methods on TCGA-GBM and CH-GBM datasets.
• Optimal performance was achieved with an image size of 190 × 215 and the Adam optimization algorithm.
• The algorithm showed a significant denoising advantage, particularly under high-intensity noise levels.

Conclusions:

• The multifeature extraction denoising algorithm effectively improves MR image resolution and reduces noise.
• The deep residual network approach, combined with multiscale feature extraction and a joint loss function, offers a robust solution for enhancing medical image quality.
• This method holds potential for improving diagnostic accuracy in clinical settings by providing clearer MR images.