An enhanced network for brain MR image denoising

Abstract

Magnetic Resonance Imaging (MRI) is a cornerstone of modern medical diagnosis due to its ability to visualize intricate soft tissues without ionizing radiation. However, noise artifacts significantly degrade image quality, hindering accurate diagnosis. Traditional denoising methods struggle to preserve details while effectively reducing noise. While deep learning approaches show promise, they often focus on local information, neglecting long-range dependencies. To address these limitations, this study proposes the deep and shallow feature fusion denoising network (DAS-FFDNet) for MRI denoising. DAS-FFDNet combines shallow and deep feature extraction with a tailored fusion module, effectively capturing both local and global image information. This approach surpasses existing methods in preserving details and reducing noise, as demonstrated on publicly available T1-weighted and T2-weighted brain image datasets. The proposed model offers a valuable tool for enhancing MRI image quality and subsequent analyses.

Keywords

Magnetic resonance imaging (MRI)image denoising deep learning UNet convolutional neural networks (CNNs)

1. Introduction

Magnetic Resonance Imaging (MRI) has become an indispensable diagnostic tool in modern medicine due to its capacity to visualize intricate soft tissue details without ionizing radiation. However, MRI images are susceptible to noise and artifacts, which can significantly impede diagnostic accuracy. Noise reduction is crucial for enhancing image clarity and contrast, enabling better visualization of anatomical structures, tumors, lesions, and blood vessels. Ultimately, this leads to more accurate diagnoses, improved treatment planning, and better patient outcomes.

Traditional image denoising techniques, such as filter-based [1,2,3] and transform-domain methods [4,5,6], often struggle to preserve image details while effectively reducing noise. Moreover, these methods require intricate parameter tuning. The advent of deep learning, particularly convolutional neural networks (CNNs), has revolutionized image processing, with models like DnCNN [7], FFDNet [8], SADNet [9], SCNN [10], and MCDnCNN [11] demonstrating promising results in MRI denoising. However, these models primarily focus on local information, limiting their ability to capture long-range dependencies inherent in MRI images.

To address this limitation, recent studies have incorporated global information into the denoising process, as seen in FFA-DMRI [12], CNN-DMRI [13], and ADNet [14]. Additionally, the encoder-decoder architecture, popularized by UNet [15], has gained traction in image denoising. While effective for conventional images, this architecture may not fully exploit the complex anatomical structures present in MRI data.

Preserving anatomical details while eliminating noise in medical images remains a critical challenge. While techniques like filtering [16,17,18], wavelet transforms [19], denoising autoencoders [20], and non-local means [21] have shown some success, they often struggle to balance detail preservation and noise reduction. CNN-based models, despite their impressive performance, face challenges such as overfitting, computational demands, and limited interpretability.

To address these limitations, this study proposes a novel MRI denoising network, deep and shallow feature fusion denoising network (DAS-FFDNet), that effectively combines the extracted deep and shallow features. The model prioritizes the preservation of both local and global image details, aiming for improved denoising performance. Key contributions include: (1) A shallow feature extraction module based on a nested double residual structure for efficient feature transmission; (2) The recursive multi-scale channel-spatial attention module (RMCSAM-UNet) aimed to capture long-range spatial information; (3) A tailored feature fusion module for MRI denoising to effectively combine shallow and deep features.

2. Related work

Zhang et al. [7] pioneered the use of deep convolutional neural networks (CNNs) for image denoising with their DnCNN model. Subsequently, the UNet architecture [15] gained prominence as a denoising baseline due to its effective multi-scale feature extraction. This encoder-decoder structure progressively downsamples and upsamples image features to capture deep receptive fields, which are then fused with high-resolution features via skip connections to mitigate information loss. Building upon this foundation, various approaches have been proposed, such as the residual dense neural network (RDUNet) [22] and the prior-based denoising network [23], which incorporate techniques like improved BM3D [24] and the non-subsampled shearlet transform (NSST-UNET) [25] to enhance denoising performance.

Denoising high-resolution brain images presents unique challenges due to their intricate anatomical structures and complex spatial relationships. Capturing long-range dependencies in these images is crucial for preserving detailed morphological information. Conventional methods often struggle to effectively address this, hindering accurate restoration. Inspired by the Noise2Noise (N2N) [26] framework, RA-UNet [27] combines the UNet with convolutional block attention module (CBAM) [28] to create a multi-scale residual block that improves noise identification and removal while preserving structural details. FONDUE [29] further advances denoising of structural MRIs by employing densely connected UNets to achieve high-quality results across different resolutions.

Recent research has explored the integration of transformers with UNet architectures to enhance image restoration. TransUNet [30] pioneers this approach, while Uformer [31] and Restormer [32] adopt transformer-based encoder-decoder structures to effectively capture both local and global image dependencies.

3. Methods

3.1. Overview of the DAS-FFDNet architecture

As illustrated in Fig. 1a, our proposed denoising network, dubbed DAS-FFDNet, comprises three primary components: deep feature extraction, shallow feature extraction, and feature fusion. The deep feature extraction module (Fig. 1b) generates high-resolution features within a UNet framework enhanced by RMCSAM-UNet [33]. The shallow feature extraction module (Fig. 1c) employs multiple dual residual blocks (DRBs) [34] to efficiently capture shallow-level image information. Finally, the adaptive fusion block combines shallow and deep features to produce the denoised image.

Figure 1.

Proposed DAS-FFDNet architecture. (a) Overview of the DAS-FFDNet architecture, comprising shallow feature extraction, RMCSAM-UNet, and feature fusion modules. (b) Detailed structure of the RMCSAM module, incorporating channel-wise and spatial-wise attention mechanisms. (c) Dual residual block (DRB) architecture employing a dual residual structure and Autocorrelation weights (ACW) for efficient feature expansion and fusion.

The shallow feature extraction module utilizes local and global skip connections to facilitate efficient feature transmission and preserve fine details. Given an input image X, the output feature $X_{s}$ of shallow feature extraction can be expressed as:

\begin{aligned} X_{S} = f_{up} (X_{n} + f_{down~} (X)) + X \end{aligned}

(1)

where

f_{u p} (\cdot)

and

f_{d o w n} (\cdot)

represent upsampling and downsampling operations, respectively.

X_{n}

is the output feature information after n DRB modules, which is expressed as:

\begin{aligned} X_{n} = f_{D R B}^{n} (X_{n - 1}) = f_{D R B}^{n} (f_{D R B}^{n - 1} (\dots (f_{D R B}^{1} (X_{0}) \dots))) \end{aligned}

(2)

where the function

f_{D R B}^{n} (\cdot)

corresponds to the operations of the n-th DRB, and

X_{0}

is the input feature of the first DRB. The output feature

Y \in R^{H \times W \times C}

can be expressed as:

\begin{aligned} Y = f_{f u s i o n} (X_{S}, X_{R}) \end{aligned}

(3)

where

X_{R}

indicates the out feature of deep feature extraction,

f_{f u s i o n} (\cdot)

is feature fusion function.

3.2. Dual residual blocks (DRBs) for shallow feature extraction

The DRB [34] comprises a nested residual structure (Fig. 1c) with an outer and inner residual unit. Each unit contains two convolutional layers with a ReLU activation. Autocorrelation weight units extract feature information after the second convolutional layer in both inner and outer units. The outer unit expands and compresses feature channels, while the inner unit focuses on feature refinement. Local and global skip connections enhance feature transmission.

The autocorrelation weight unit employs global average pooling followed by a Sigmoid function to generate weights for feature fusion. This mechanism stabilizes the model and improves feature learning.

3.3. RMCSAM-UNet for deep feature extraction

The RMCSAM [33] captures multi-scale attention information in feature maps, enhancing long-range dependencies. We integrate RMCSAM into a UNet architecture to form the RMCSAM-UNet module (Fig. 1a). The encoder downsamples features and the decoder upsamples them, with RMCSAM modules at each level to capture multi-scale information. Skip connections transfer features between encoder and decoder for effective fusion.

3.4. Feature fusion module

To combine shallow and deep features, we employ an adaptive fusion block. The shallow features are upsampled and concatenated with deep features. Subsequent convolutional and deformable convolutional layers [35] process the combined features. Adaptive fusion is achieved using element-wise multiplication and skip connections. The output feature $Y$ can be expressed as:

\begin{matrix} Y = W_{2} σ (W_{1} X_{S R}) \end{matrix}

(4)

\begin{matrix} X_{S R} = f_{d e f o r m} (W_{1} f_{c o n c a t} (f_{t r a n s} (X_{S}), X_{R})) + f_{t r a n s} (X_{S}) \end{matrix}

(5)

where

f_{t r a n s} (\cdot)

represents the transpose convolution operation,

f_{c o n c a t} (\cdot)

denotes the concatenation operation,

W_{1}

and

W_{2}

stand for the

3 \times 3

convolutional layers,

f_{d e f o r m} (\cdot)

signifies the deformable convolution, and

σ (\cdot)

corresponds to the ReLU activation function.

4. Experiments

4.1. Datasets

We evaluated our model using three publicly available MRI datasets: the IXI dataset (https://brain-development.org/ixi-dataset/), the MICCAI 2013 grand challenge open dataset (https://www.synapse.org/#!Synapse:syn3193805/wiki/217780), and a dataset from Karolinska University Hospital (KUH) with varying levels of acceleration and noise [36].

The IXI dataset includes three types of MRI contrasts–T1-weighted (T1w), T2-weighted (T2w), and Proton Density (PD) weighted images – acquired at three different sites using 1.5T or 3T scanners. The MICCAI dataset contains only T1w brain images. The KUH dataset comprises T1w images with two levels of signal-to-noise ratios (SNR), acquired at 3T using MPRAGE (magnetization-prepared rapid acquisition with gradient-echo) and wave-CAIPI (Controlled Aliasing In Parallel Imaging) MPRAGE sequences. MPRAGE is the gold-standard sequence for assessing brain anatomy and is generally preferred for voxel-based morphometric (VBM) analyses. Conventional MPRAGE sequences were acquired with a 2-fold GRAPPA accelerated protocol, taking 4:26 minutes to acquire a whole-brain 3D T1w volume with 1 mm isotropic spatial resolution. The 9-fold accelerated wave-CAIPI MPRAGE protocol took only 2 minutes to acquire a whole-brain 3D T1w volume with the same spatial resolution.

From the IXI (T1w and T2w) and MICCAI (T1w) datasets, we randomly selected 80 volumes of 3D MRI brain images per dataset, splitting them into training, validation, and test sets with an 8:1:1 ratio. The central slab of each volume was cropped to 100 slices of 2D images, each 256 $\times$ 256 pixels in size. For model training, we used these preprocessed images as ground truth reference images and introduced Rician noise at various levels (3%, 5%, 7%, 9%) to create training data pairs.

The KUH dataset contains T1w image volumes at two different SNR levels. We selected 48 volumes of 3D T1w MRI brain images at each SNR level, cropping the central slab of each volume to 120 slices of 2D images, each 256 $\times$ 256 pixels in size. Similar to the other datasets, these volumes were split into training, validation, and test sets with a 38:5:5 ratio. To investigate how the noise level in the MRI data affects the image segmentation results, we performed VBM analysis of the KUH dataset which containing dynamic series of 20 timeframes of 2x-GRAPPA and 9x-Wave-CAIPI T1w images. The T1w image data were pre-processed with the CAT12 Toolbox (http://www.neuro.uni-jena.de/cat/) in Statistical Parametric Mapping (SPM12) (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). The processing pipeline was used to segment the images into grey matter (GM), white matter (WM) and cerebrospinal fluid probability maps registered to the Montreal Neurological Institute (MNI) standard space. To investigate the difference in the segmented tissue volume between the two protocol variants is related to the image noise levels, we calculated also the mean, standard deviation (Sd), and coefficient of variance (CVAR) for the segmented images.

4.2. Implementation details

Experiments were conducted on an HP Z840 Linux workstation (Ubuntu 20.04) equipped with dual Intel Xeon E5-2687W V4 CPUs, 512 GB DDR4-2400 ECC registered SDRAM, and an NVIDIA Tesla V100 PCIe GPU with 16 GB RAM. The model was implemented using PyTorch 1.11.0 and leveraged the CUDA 11.3 platform for GPU acceleration. For training, the ADAM optimizer [37] was employed with a learning rate of 1e-4 and momentum decay of 1e-8. The model was trained for 50 epochs using a batch size of 4. The proposed model was benchmarked against several state-of-the-art denoising methods, including DnCNN [7], FFDNet [8], CNN-DMRI [13], SADNet [9], RDUNet [22], MSANet [35] and LADCNN [38].

4.3. Metrics for performance evaluation

To assess model performance, peak signal-to-noise ratio (PSNR) [39] and structural similarity index metrics (SSIM) [40] were computed. PSNR is a metric used to assess image quality by comparing a reconstructed image to its original counterpart. It is expressed in decibels (dB), with higher values indicating better image quality. PSNR is calculated by determining the mean squared error between the images and then applying a logarithmic transformation. While simple to compute, PSNR has limitations as it is highly sensitive to noise and often poorly correlates with human perception of image quality. SSIM is a perceptual metric designed to measure the structural similarity between two images. It incorporates luminance, contrast, and structure information to provide a more comprehensive assessment of image quality. SSIM values range from 0 to 1, with values closer to 1 indicating higher similarity. Unlike PSNR, SSIM better aligns with human perception and is more robust to noise. Consequently, SSIM is generally preferred over PSNR for evaluating image quality.

5. Results

The main experimental results are summarized in Table 1. To facilitate comparison, the PSNR and SSIM metrics after the various denoising methods are presented in Fig. 2. As expected, image quality deteriorates with increasing noise levels, regardless of the denoising approach. The PSNR decreased by approximately 5 dB when the noise level tripled. The PSNR performance of the different denoising algorithms varied by about 3–4 dB. Overall, the proposed DAS-FFDNet demonstrated superior performance across all datasets. Its advantage was particularly evident in the higher noise levels (7 and 9%). The RDUnet showed good performance in the lower noise levels (3 and 5%). Our model exhibited remarkable robustness and excelled at preserving spatial information in the face of increasing noise. This highlights the method’s precision in retaining intricate brain details, effectively restoring both morphology and anatomical structures.

Table 1
Denoising performance comparison of the proposed RMCSAM-UNet with state-of-the-art methods in terms of PSNR (dB) and SSIM on the IXI and MICCAI datasets. The best results in each comparison group are highlighted in bold.

Noise level Model IXI(T1w) IXI(T2w) MICCAI(T1w)

PSNR/SSIM PSNR/SSIM PSNR/SSIM

3% DnCNN [7] 40.32/0.9885 42.14/0.9923 42.88/0.9950

FFDNet [8] 38.34/0.9861 41.04/0.9919 43.52/0.9956

CNNDMRI [13] 37.79/0.9824 38.94/0.9851 40.03/0.9894

SADNet [9] 38.00/0.9815 38.65/0.9739 40.22/0.9814

RDUNet [22] 40.64/0.9913 42.10/0.9928 44.17/0.9955

MSANet [35] 39.48/0.9832 40.77/0.9842 42.77/0.9899

LADCNN [38] 40.06/0.9882 41.88/0.9888 43.99/0.9937

Ours 40.12/0.9888 42.18/0.9917 43.89/0.9957

5% DnCNN [7] 37.38/0.9806 39.33/0.9864 40.83/0.9882

FFDNet [8] 37.13/0.9816 38.96/0.9880 40.20/0.9898

CNNDMRI [13] 36.13/0.9727 37.55/0.9831 39.97/0.9765

SADNet [9] 36.69/0.9783 36.74/0.9612 38.82/0.9734

RDUNet [22] 37.79/0.9845 39.39/0.9885 41.25/0.9917

MSANet [35] 36.37/0.9710 37.85/0.9737 39.69/0.9827

LADCNN [38] 37.45/0.9832 39.38/0.9848 40.69/0.9889

Ours 37.85/0.9832 39.67/0.9886 41.21/0.9902

7% DnCNN [7] 36.05/0.9780 37.75/0.9827 38.85/0.9813

FFDNet [8] 35.55/0.9697 37.13/0.9825 39.45/0.9891

CNNDMRI [13] 35.26/0.9721 37.24/0.9780 38.74/0.9703

SADNet [9] 35.79/0.9626 35.46/0.9523 37.28/0.9671

RDUNet [22] 35.82/0.9773 37.04/0.9820 39.56/0.9892

MSANet [35] 34.97/0.9677 36.47/0.9742 37.97/0.9733

LADCNN [38] 35.74/0.9769 36.96/0.9839 38.93/0.9845

Ours 36.26/0.9779 37.93/0.9859 39.70/0.9898

9% DnCNN [7] 34.78/0.9712 36.26/0.9581 37.49/0.9443

FFDNet [8] 34.32/0.9674 35.97/0.9790 38.23/0.9854

CNNDMRI [13] 33.96/0.9649 35.58/0.9723 35.69/0.9626

SADNet [9] 33.92/0.9440 34.54/0.9519 36.47/0.9612

RDUNet [22] 34.25/0.9679 35.48/0.9757 37.97/0.9835

MSANet [35] 33.15/0.9581 35.14/0.9680 36.66/0.9704

LADCNN [38] 34.03/0.9641 36.26/0.9798 36.93/0.9754

Ours 35.30/0.9745 37.02/0.9802 38.82/0.9866

Noise level	Model	IXI(T1w)	IXI(T2w)	MICCAI(T1w)
3%	DnCNN [7]	40.32/0.9885	42.14/0.9923	42.88/0.9950
	FFDNet [8]	38.34/0.9861	41.04/0.9919	43.52/0.9956
	CNNDMRI [13]	37.79/0.9824	38.94/0.9851	40.03/0.9894
	SADNet [9]	38.00/0.9815	38.65/0.9739	40.22/0.9814
	RDUNet [22]	40.64/0.9913	42.10/0.9928	44.17/0.9955
	MSANet [35]	39.48/0.9832	40.77/0.9842	42.77/0.9899
	LADCNN [38]	40.06/0.9882	41.88/0.9888	43.99/0.9937
	Ours	40.12/0.9888	42.18/0.9917	43.89/0.9957
5%	DnCNN [7]	37.38/0.9806	39.33/0.9864	40.83/0.9882
	FFDNet [8]	37.13/0.9816	38.96/0.9880	40.20/0.9898
	CNNDMRI [13]	36.13/0.9727	37.55/0.9831	39.97/0.9765
	SADNet [9]	36.69/0.9783	36.74/0.9612	38.82/0.9734
	RDUNet [22]	37.79/0.9845	39.39/0.9885	41.25/0.9917
	MSANet [35]	36.37/0.9710	37.85/0.9737	39.69/0.9827
	LADCNN [38]	37.45/0.9832	39.38/0.9848	40.69/0.9889
	Ours	37.85/0.9832	39.67/0.9886	41.21/0.9902
7%	DnCNN [7]	36.05/0.9780	37.75/0.9827	38.85/0.9813
	FFDNet [8]	35.55/0.9697	37.13/0.9825	39.45/0.9891
	CNNDMRI [13]	35.26/0.9721	37.24/0.9780	38.74/0.9703
	SADNet [9]	35.79/0.9626	35.46/0.9523	37.28/0.9671
	RDUNet [22]	35.82/0.9773	37.04/0.9820	39.56/0.9892
	MSANet [35]	34.97/0.9677	36.47/0.9742	37.97/0.9733
	LADCNN [38]	35.74/0.9769	36.96/0.9839	38.93/0.9845
	Ours	36.26/0.9779	37.93/0.9859	39.70/0.9898
9%	DnCNN [7]	34.78/0.9712	36.26/0.9581	37.49/0.9443
	FFDNet [8]	34.32/0.9674	35.97/0.9790	38.23/0.9854
	CNNDMRI [13]	33.96/0.9649	35.58/0.9723	35.69/0.9626
	SADNet [9]	33.92/0.9440	34.54/0.9519	36.47/0.9612
	RDUNet [22]	34.25/0.9679	35.48/0.9757	37.97/0.9835
	MSANet [35]	33.15/0.9581	35.14/0.9680	36.66/0.9704
	LADCNN [38]	34.03/0.9641	36.26/0.9798	36.93/0.9754
	Ours	35.30/0.9745	37.02/0.9802	38.82/0.9866

Figure 2.

Denoising performance evaluation of the proposed RMCSAM-UNet in comparison to state-of-the-art methods. PSNR and SSIM metrics are plotted as a function of noise level for IXI (T1w, T2w) and MICCAI (T1w) datasets to assess image quality preservation.

Figure 3.

Visual comparison of denoising performance between the proposed DAS-FFDNet and state-of-the-art methods on IXI (T1w, T2w) and MICCAI (T1w) datasets with 9% added noise. Representative sagittal, axial, and coronal slices are shown for each dataset, along with zoomed-in regions (marked by red rectangles) and difference images between denoised and ground truth images.

Figure 3 depicts representative denoised images produced by the different methods, accompanied by zoomed-in views. For visual comparison, discrepancies between the original and denoised images are also shown. The denoised image comparisons confirm the superior ability of our proposed method to preserve fine brain details while effectively removing noise artifacts compared to other approaches. To investigate the impact of the number of DRB blocks on model performance, we systematically varied the number of DRBs in the DAS-FFDNet from 0 to 4 and evaluated the results. The ablation study using the IXI(T2w) dataset across different noise levels is summarized in Table 2. The results indicate that increasing the number of DRB modules generally improves PSNR and SSIM performance. As shown in Fig. 4, adding a single DRB module initially led to a slight decline in PSNR and SSIM compared to the baseline. However, as the number of DRB modules increased, performance recovered and continued to improve, highlighting the need to balance performance and model complexity. At last, we select 3 DRBs in our proposed method.

Table 2

PSNR (dB)/SSIM performances of ablation studies on the IXI(T2w) dataset.

Modules	Params	3%	5%	7%	9%
Baseline	4.93M	40.06/0.9861	38.18/0.9850	37.54/0.9779	36.27/0.9760
baseline $+$ 1DRB	5.02M	38.85/0.9841	37.89/0.9822	36.60/0.9685	35.48/0.9607
baseline $+$ 2DRB	5.11M	40.80/0.9871	38.28/0.9833	37.33/0.9815	36.45/0.9769
baseline $+$ 3DRB	5.20M	41.51/0.9885	39.13/0.9879	37.70/0.9819	36.48/0.9770
baseline $+$ 4DRB	5.29M	41.87/0.9893	39.52/0.9867	37.92/0.9849	36.73/0.9780

Table 3

Summary of the VBM results for the two serial datasets acquired from a single subject using the two protocol variants to assess the effect of noise levels. The overlapping voxels are voxels with partially overlapping GM and WM tissues.

Dataset	PSNR (dB)	SSIM	GM (%)	WM (%)	Overlapping (%)
2x-GRAPP	30.21	0.9267	43.75 $\pm$ 0.11	35.80 $\pm$ 0.13	28.43 $\pm$ 0.11
9x-wave CAIP	25.23	0.8177	43.28 $\pm$ 0.18	36.07 $\pm$ 0.18	28.10 $\pm$ 0.19

Figure 4.

Bar graphs PSNR (a) and SSIM (b) metrics for the IXI (T2w) dataset after denoising with DAS-FFDnet models with different number of DRBs.

To assess how noise levels in MRI data affect image segmentation, we compared VBM results for a time-series T1w dataset from KUH acquired using 2x GRAPPA and 9x Wave-CAIPI acceleration techniques. Figure 5 displays cross-sectional views of a typical 1 mm isotropic resolution T1w volume acquired using both protocols. As expected, noise levels differed significantly between the time-series datasets. This is further supported by the statistical analysis of the corresponding VBM results shown in Fig. 6, which depicts cross-sectional views of mean GM concentration (a and b), SD (c and d), and CVAR (e and f) for the VBM serial datasets derived from 20 repeated MPRAGE measurements in a single subject. The VBM data acquired with the 9-fold Wave-CAIPI encoding protocol exhibited increased noise levels, reflected in more voxels with higher SD and CVAR (see Fig. 6c-f), particularly in the central brain region and at the boundaries between GM and WM tissues. Interestingly, some frontal regions showed lower noise levels compared to the 2-fold GRAPPA scans. Table 3 summarizes the VBM results for both serial datasets using the two protocols. After denoising with the proposed DAS-FFDNet framework, the PSNR of T1w images acquired with 9x Wave-CAIPI was approximately 5 dB lower than those acquired with 2x GRAPPA, and the corresponding SSIM was 11.8% lower. It is clear that the reduced image quality associated with faster acquisition impacted segmentation results. As shown in Table 3, VBM analysis for the 9-fold Wave-CAIPI MPRAGE protocol yielded 0.47% less GM and 0.13% more WM compared to the 2-fold GRAPPA protocol. Additionally, the VBM results for the 2-fold GRAPPA scans detected 0.33% more voxels with overlapping GM and WM tissues.

Figure 5.

Cross-sectional displays of a typical T1w 3D MRI volume for a healthy adult subject at 1 mm isotropic resolution acquired at 3T using 2x-GRAPPA (a) and 9x-Wave CAIPI acceleration methods.

Figure 6.

Cross-sectional displays of the mean GM concentration (a and b), standard deviation (c and d), and coefficient of variance (e and f) of the VBM serial datasets derived from the 20 timeframes of repeated measurements from a single subject. The left (a, c, and e) and right (b, d, and f) panels are the corresponding statistical maps (mean, SD, and CVAR) derived from the 2x-GRAPPA and 9x-Wave CAIPI protocols, respectively. The crossing green lines indicate the location of the cross sections.

6. Discussion

6.1. The impact of noise on MRI image quality and subsequent analysis

Noise, a ubiquitous artifact in MRI imaging, significantly degrades image quality, compromising diagnostic accuracy and subsequent image analysis tasks. The presence of noise reduces image contrast, obscures fine details, and introduces artifacts that can mislead interpretation. This study underscores the critical role of noise reduction in optimizing MRI image quality and subsequent downstream analyses, such as segmentation.

While the proposed DAS-FFDNet effectively mitigates noise and preserves image details, its limitations become apparent in scenarios with extremely high noise levels, often introduced by accelerated acquisition techniques. As demonstrated by VBM analysis, noise-induced discrepancies in brain tissue volume measurements highlight the need for robust denoising techniques to ensure accurate neuroimaging analysis.

6.2. The DAS-FFDNet: A Novel approach to image denoising

The DAS-FFDNet represents a significant advancement in MRI image denoising through its innovative combination of shallow and deep feature extraction. This approach effectively addresses the complex challenge of preserving both fine-grained local details and essential global image structures, which are crucial for accurate clinical diagnosis. By incorporating DRBs within the shallow feature extraction module, the model excels in preserving image details, particularly in challenging high-noise scenarios.

Comparative analysis with state-of-the-art CNN-based denoising methods (DnCNN, FFDNet, CNN-DMRI, SADNet, RDUNet, MSANet and LADCNN) demonstrates the DAS-FFDNet’s superior performance in terms of noise reduction and image detail preservation. The model’s robustness and adaptability to various image content make it a promising tool for a wide range of clinical applications.

6.3. Challenges and future directions in denoising MRI data of different contrasts

While this study focused on T1w and T2w brain images, the complexities inherent to different MRI contrasts and anatomical regions necessitate tailored denoising approaches. For instance, PD-weighted images, often used for visualizing knee ligaments, require distinct denoising strategies compared to brain imaging. The specific diagnostic targets and underlying image characteristics of each modality influence the optimal denoising techniques. Future research should explore the development of adaptive denoising methods that can be tailored to different MRI contrasts and anatomical regions. This would involve considering factors such as tissue contrast, noise patterns, and specific diagnostic requirements to optimize denoising performance.

6.4. Limitations and future directions in MRI image denoising

While the DAS-FFDNet exhibits promising results, further research is necessary to fully realize its potential. Key areas for exploration include: (1) Modality expansion: Extending the model’s applicability to other MRI modalities (DWI, fMRI) to address the unique challenges posed by these imaging techniques. (2) Generative adversarial networks (GANs) integration: Incorporating GANs to enhance image realism and quality, potentially leading to improved perceptual image quality. (3) Advanced evaluation metrics: Developing more sophisticated evaluation metrics that correlate closely with human perception to accurately assess the model’s impact on downstream tasks. (4) Real-time implementation: Investigating real-time implementations for clinical adoption and integration into clinical workflows. By addressing these challenges, the DAS-FFDNet can be further refined to become an indispensable tool for improving diagnostic accuracy and patient care.

7. Conclusion

The DAS-FFDNet represents a significant advancement in MRI image denoising, offering a robust and effective solution for preserving image quality while reducing noise. Its superior performance compared to state-of-the-art methods underscores its potential to improve medical image analysis. However, the diverse nature of MRI contrasts and anatomical regions necessitates further research to develop tailored denoising approaches. By addressing these challenges and exploring future research directions, the model’s impact on clinical practice can be maximized.

Footnotes

Acknowledgments

This research was supported by a grant from the Zhejiang Natural Science Foundation of China (No. LY23F010005). Our study utilized three publicly available magnetic resonance imaging (MRI) datasets: the IXI dataset (https://brain-development.org/ixi-dataset/), the MICCAI 2013 grand challenge open dataset (https://www.synapse.org/# !Synapse:syn3193805/wiki/217780), and a dataset from Karolinska University Hospital [].

References

Gerig

et al., Nonlinear anisotropic filtering of MRI data, IEEE Transations on Medical Imaging 11 (1992), 221–232.

Wong

W.C.K.

Chung

A.C.S.

S.C.H.

, Trilateral filtering for biomedical images, in: 2004 2nd IEEE International Symposium on Biomedical Imaging: Macro to Nano (IEEE Cat No. 04EX821), IEEE, Arlington, VA, USA, 2004, pp. 820–823.

Maneesha Mohan

M.R.

Sulochana

C.H.

Latha

, Medical image denoising using multistage directional median filter, in: 2015 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2015], IEEE, Nagercoil, 2015, pp. 1–6.

Satheesh

Prasad

, Medical Image Denoising Using Adaptive Threshold Based on Contourlet Transform, Advanced Computing: An International Journal 2 (2011), 52–58.

Vanitha

Rajeswari

Ebenezer

, An Analysis and Reduction of Fractional Brownian motion Noise in Biomedical Images Using Curvelet Transform and Various Filtering and Thresholding Techniques, International Journal of Engineering Science 5(10) (2016), 18–27.

Biswas

Purkayastha

Roy

, Denoising of MRI Images Using Curvelet Transform, in: Advances in Systems, Control and Automation Konkani

Bera

Paul

, eds, Springer Singapore, Singapore, 2018, pp. 575–583.

Zhang

et al., Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising, IEEE Transactions on Image Processing 26 (2017), 3142–3155.

Zhang

Zuo

Zhang

, FFDNet: Toward a Fast and Flexible Solution for CNN based Image Denoising, IEEE Transactions on Image Processing 27 (2018), 4608–4622.

Chang

et al., Spatial-Adaptive Network for Single Image Denoising, in: Computer Vision – ECCV 2020, Vedaldi

Bischof

Brox

Frahm

JM.

, eds., Lecture Notes in Computer Science, Springer, Cham, 2020, pp. 171–187.

10.

Isogawa

et al., Deep Shrinkage Convolutional Neural Network for Adaptive Noise Reduction, IEEE Signal Processing Letters 25 (2018), 224–228.

11.

Jiang

et al., Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network, Japanese Journal of Radiology 36 (2018), 566–574.

12.

Hong

et al., FFA-DMRI: A Network Based on Feature Fusion and Attention Mechanism for Brain MRI Denoising, Frontiers in Neuroscience 14 (2020), 577–937.

13.

Tripathi

P.C.

Bag

, CNN-DMRI: A Convolutional Neural Network for Denoising of Magnetic Resonance Images, Pattern Recognition Letters 135 (2020), 57–63.

14.

Tian

et al., Attention-guided CNN for image denoising, Neural Networks 124 (2020), 117–129.

15.

Ronneberger

Fischer

Brox

, U-Net: Convolutional Networks for Biomedical Image Segmentation, in: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 N. Navab, Hornegger

Wells

Frangi

, eds, Lecture Notes in Computer Science, 2015, pp. 234–241.

16.

et al., A Numerical-Integration Perspective on Gaussian Filters, IEEE Transactions on Signal Processing 54 (2006), 2910–2921.

17.

Tang

et al., An Adaptive Anisotropic Diffusion Filter for Noise Reduction in MR Images, in: 2007 International Conference on Mechatronics and Automation, IEEE, Harbin, China, 2007, pp. 1299–1304.

18.

Hamarneh

Hradsky

, Bilateral Filtering of Diffusion Tensor Magnetic Resonance Images, IEEE Transacttions on Image Processing 16 (2007), 2463–2475.

19.

Phatak

et al., Wavelet domain filtering of MR image sequences with appropriate filtering approach based on MR image type, in: 2011 International Conference on Signal Processing, Communication, Computing and Networking Technologies, IEEE, Thuckalay, Tamil Nadu, India, 2011, pp. 520–525.

20.

Gondara

, Medical image denoising using convolutional denoising autoencoders, in: 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW), 2016, pp. 241–246.

21.

Bhujle

H.V.

Vadavadagi

B.H.

, NLM based magnetic resonance image denoising – A review, Biomedical Signal Processing and Control 47 (2019), 252–261.

22.

Gurrola-Ramos

Dalmau

Alarcon

T.E.

, A Residual Dense U-Net Neural Network for Image Denoising, IEEE Access 9 (2021) 31742–31754.

23.

Aetesam

Maji

S.K.

, Attention-Based Noise Prior Network for Magnetic Resonance Image Denoising, in: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), IEEE, Kolkata, India, 2022, pp. 1–4.

24.

Dabov

et al., Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering, IEEE Transactions on Image Processing 16 (2007), 2080–2095.

25.

et al., GPR image denoising with NSST-UNET and an improved BM3D, Digital Signal Processing 123 (2022), 103–402.

26.

Lehtinen

et al., Noise2Noise: Learning Image Restoration without Clean Data, Arxiv (2018).

27.

Liu

, and Huang

, RA-UNet: an improved network model for image denoising, The Visual Computer 40 (2023), 4319–4335.

28.

Woo

et al., CBAM: Convolutional Block Attention Module, in: Computer Vision – ECCV 2018, Ferrari

Hebert

Sminchisescu

Weiss

, eds., Springer International Publishing, Cham, 2018, pp. 3–19.

29.

Adame-Gonzalez

et al., FONDUE: Robust resolution-invariant denoising of MR Images using Nested UNets, Neuroscience (2023).

30.

Chen

et al., TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation, Arxiv, abs/210204306. (2021).

31.

Wang

et al., Uformer: A General U-Shaped Transformer for Image Restoration, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New Orleans, LA, USA, 2022, pp. 17662–17672.

32.

Zamir

S.W.

et al., Restormer: Efficient Transformer for High-Resolution Image Restoration, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New Orleans, LA, USA, 2022, pp. 5718–5729.

33.

Liu

et al., Recursive Multi-Scale Channel-Spatial Attention for Fine-Grained Image Classification, IEICE Transactions on Information and Systems 105–D (2022), 713–726.

34.

Chen

et al., MFFN: image super-resolution via multi-level features fusion network, The Visual Computer 40 (2024), 489–504.

35.

Gou

et al., Multi-Scale Adaptive Network for Single Image Denoising, in: Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS ’22) Curran Associates Inc., Red Hook, NY, USA, 2024, pp. 14099–14112.

36.

Olivo

et al., Estimated gray matter volume rapidly changes after a short motor task, Cerebral Cortex 32 (2022), 4356–4369.

37.

Kingma

D.P.

, Adam: A Method for Stochastic Optimization, arXiv preprint arXiv:14126980. (2014).

38.

Shi

Liu

, Ultrasound image denoising autoencoder model based on lightweight attention mechanism, Quantitative Imaging in Medicine and Surgery 14 (2024), 3557–3571.

39.

Chan

L.C.

Whiteman

, Hardware-constrained hybrid coding of video imagery, IEEE Transactions on Aerospace and Electronic Systems 1 (1983), 71–84.

40.

Wang

et al., Image Quality Assessment: From Error Visibility to Structural Similarity, IEEE Transactions on Image Processing 13 (2004), 600–612.

Noise level	Model	IXI(T1w)	IXI(T2w)	MICCAI(T1w)
		PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
3%	DnCNN [7]	40.32/0.9885	42.14/0.9923	42.88/0.9950
	FFDNet [8]	38.34/0.9861	41.04/0.9919	43.52/0.9956
	CNNDMRI [13]	37.79/0.9824	38.94/0.9851	40.03/0.9894
	SADNet [9]	38.00/0.9815	38.65/0.9739	40.22/0.9814
	RDUNet [22]	40.64/0.9913	42.10/0.9928	44.17/0.9955
	MSANet [35]	39.48/0.9832	40.77/0.9842	42.77/0.9899
	LADCNN [38]	40.06/0.9882	41.88/0.9888	43.99/0.9937
	Ours	40.12/0.9888	42.18/0.9917	43.89/0.9957
5%	DnCNN [7]	37.38/0.9806	39.33/0.9864	40.83/0.9882
	FFDNet [8]	37.13/0.9816	38.96/0.9880	40.20/0.9898
	CNNDMRI [13]	36.13/0.9727	37.55/0.9831	39.97/0.9765
	SADNet [9]	36.69/0.9783	36.74/0.9612	38.82/0.9734
	RDUNet [22]	37.79/0.9845	39.39/0.9885	41.25/0.9917
	MSANet [35]	36.37/0.9710	37.85/0.9737	39.69/0.9827
	LADCNN [38]	37.45/0.9832	39.38/0.9848	40.69/0.9889
	Ours	37.85/0.9832	39.67/0.9886	41.21/0.9902
7%	DnCNN [7]	36.05/0.9780	37.75/0.9827	38.85/0.9813
	FFDNet [8]	35.55/0.9697	37.13/0.9825	39.45/0.9891
	CNNDMRI [13]	35.26/0.9721	37.24/0.9780	38.74/0.9703
	SADNet [9]	35.79/0.9626	35.46/0.9523	37.28/0.9671
	RDUNet [22]	35.82/0.9773	37.04/0.9820	39.56/0.9892
	MSANet [35]	34.97/0.9677	36.47/0.9742	37.97/0.9733
	LADCNN [38]	35.74/0.9769	36.96/0.9839	38.93/0.9845
	Ours	36.26/0.9779	37.93/0.9859	39.70/0.9898
9%	DnCNN [7]	34.78/0.9712	36.26/0.9581	37.49/0.9443
	FFDNet [8]	34.32/0.9674	35.97/0.9790	38.23/0.9854
	CNNDMRI [13]	33.96/0.9649	35.58/0.9723	35.69/0.9626
	SADNet [9]	33.92/0.9440	34.54/0.9519	36.47/0.9612
	RDUNet [22]	34.25/0.9679	35.48/0.9757	37.97/0.9835
	MSANet [35]	33.15/0.9581	35.14/0.9680	36.66/0.9704
	LADCNN [38]	34.03/0.9641	36.26/0.9798	36.93/0.9754
	Ours	35.30/0.9745	37.02/0.9802	38.82/0.9866