FFA-DMRI: A Network Based on Feature Fusion and Attention Mechanism for Brain MRI Denoising

Danfeng Hong, Chenxi Huang, Chenhui Yang, Jianpeng Li, Yunhan Qian, Chunting CaiFrontiers in Neuroscience 2020 · DOI

Abstract

Magnetic Resonance Imaging (MRI) is an indispensable tool in the diagnosis of brain diseases due to painlessness and safety. Nevertheless, Rician noise is inevitably injected during the image acquisition process, which leads to poor observation and interferes with the treatment. Owing to the complexity of Rician noise, using the elimination method of Gaussian to remove it does not perform well. Therefore, the feature fusion and attention network (FFA-DMRI) is proposed to separate noise from observed MRI. Inspired by the attention-guided CNN network (ADNet) and Convolutional block attention module (CBAM), a spatial attention mechanism has been specially designed to obtain the area of interest in MRI. Furthermore, the feature fusion block concatenates local with global information, which makes full use of the multilevel structure and boosts the expressive ability of network. The comprehensive experiments on Alzheimer's disease neuroimaging initiative dataset (ADNI) have demonstrated high effectiveness of FFA-DMRI with maintaining the crucial brain details. Moreover, in terms of visual inspections, the denoising results are also consistent with human perception.

Introduction

Magnetic Resonance Imaging (MRI) is a cornerstone of modern neurology, but the images it produces are often compromised by Rician noise introduced during acquisition. Unlike standard Gaussian noise, Rician noise is signal-dependent, meaning its intensity varies with the brightness of the underlying tissue. This characteristic makes it particularly difficult to remove using traditional filters, which often over-smooth images and blur critical anatomical boundaries, such as those between gray and white matter.

The FFA-DMRI Architecture

To address these challenges, researchers developed the Feature Fusion and Attention Network (FFA-DMRI). The model is designed to learn the residual noise (the noise itself) rather than the clean image directly. By predicting the noise and subtracting it from the original noisy input, the model achieves more stable training and better preservation of structural details.

Figure 1: Overall architecture of the proposed FFA-DMRI network, highlighting the feature extraction, fusion, and attention blocks.

The network architecture consists of three specialized stages:

  1. Feature Extraction Block: Utilizes a combination of standard and dilated convolutions. Dilated convolutions expand the receptive field, allowing the network to capture broader contextual information without increasing parameter counts or sacrificing resolution.
  2. Feature Fusion Block: To prevent the loss of fine details in deeper layers, this block concatenates intermediate feature maps with the original noisy input. This fusion ensures that low-level structural information is preserved throughout the reconstruction.
  3. Attention Block: Incorporates both spatial and channel attention mechanisms to focus the network's resources on informative regions and suppress noise-heavy channels.

Optimized Spatial Attention

A key innovation in FFA-DMRI is the optimization of the spatial attention module. While many attention modules use average pooling, the authors found that maximum pooling is superior for MRI data.

Figure 2: The spatial attention module utilizing maximum pooling to better highlight brain textures against dark backgrounds.

Because MRI backgrounds are predominantly black (values near zero), average pooling can dilute the signal of interest. Max pooling effectively "highlights" the brain textures, helping the attention mechanism lock onto the areas of interest. Additionally, the authors intentionally omitted Batch Normalization (BN) in the initial layers to preserve the original image contrast, which is vital for distinguishing signal-dependent Rician noise from real tissue.

Experimental Results

The model was evaluated using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, with synthetic Rician noise added at various intensity levels ($\lambda$ = 5, 10, 20, 30).

FFA-DMRI was compared against traditional filters (BM3D, NLM) and state-of-the-art deep learning models (DnCNN, IRCNN). Quantitatively, FFA-DMRI achieved the highest scores in both PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) across all noise levels. For instance, at a noise level of 20, it outperformed the popular IRCNN by 1.43 dB in PSNR and improved SSIM by 2.66%.

Qualitatively, visual inspections confirmed that FFA-DMRI maintains sharp boundaries and clear contrast between brain regions, avoiding the blurring or mottled artifacts common in other denoising methods.

Conclusion

FFA-DMRI represents a significant advancement in MRI denoising by effectively combining dilated convolutions, feature fusion, and a domain-specific attention mechanism. By prioritizing the preservation of structural details and optimizing for the unique characteristics of Rician noise, the network provides high-quality images that are consistent with human perception and clinical requirements. Future work will focus on validating the model on real-world noisy scans and exploring its performance in "blind" denoising scenarios.

Summary updated

Abstract

Magnetic Resonance Imaging (MRI) is an indispensable tool in the diagnosis of brain diseases due to painlessness and safety. Nevertheless, Rician noise is inevitably injected during the image acquisition process, which leads to poor observation and interferes with the treatment. Owing to the complexity of Rician noise, using the elimination method of Gaussian to remove it does not perform well. Therefore, the feature fusion and attention network (FFA-DMRI) is proposed to separate noise from observed MRI. Inspired by the attention-guided CNN network (ADNet) and Convolutional block attention module (CBAM), a spatial attention mechanism has been specially designed to obtain the area of interest in MRI. Furthermore, the feature fusion block concatenates local with global information, which makes full use of the multilevel structure and boosts the expressive ability of network. The comprehensive experiments on Alzheimer's disease neuroimaging initiative dataset (ADNI) have demonstrated high effectiveness of FFA-DMRI with maintaining the crucial brain details. Moreover, in terms of visual inspections, the denoising results are also consistent with human perception.
BibTeX
@inproceedings{hong2020ffa,
  author = {Danfeng Hong and Chenxi Huang and Chenhui Yang and Jianpeng Li and Yunhan Qian and Chunting Cai},
  title = {{FFA-DMRI: A Network Based on Feature Fusion and Attention Mechanism for Brain MRI Denoising}},
  booktitle = {Frontiers in Neuroscience},
  year = {2020},
  doi = {10.3389/fnins.2020.577937},
}
Figure