See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358682705

ADMM based Deep Denoiser Prior for Enhancing Single Coil Magnitude MR
images

Conference Paper · December 2021

DOI: 10.1109/ACTS53447.2021.9708111

CITATION READS
1 30

3 authors, including:

Hari Kishan Repala

National Institute of Technology Calicut
3 PUBLICATIONS 3 CITATIONS

SEE PROFILE

All content following this page was uploaded by Hari Kishan Repala on 04 November 2022.

The user has requested enhancement of the downloaded file.

 ADMM based Deep Denoiser Prior for Enhancing
Single Coil Magnitude MR images
Aneeta Christopher Harikishan R
2021 Advanced Communication Technologies and Signal Processing (ACTS) | 978-1-6654-2337-3/21/$31.00 ©2021 IEEE | DOI: 10.1109/ACTS53447.2021.9708111

Dept. of Electronics and Communication Engineering Dept. of Electronics and Communication Engineering
National Institute of Technology Calicut National Institute of Technology Calicut
Kozhikode, Kerala, India Kozhikode, Kerala, India
[email protected] harikishan [email protected]

Sudeep P V*
Dept. of Electronics and Communication Engineering
National Institute of Technology Calicut
Kozhikode, Kerala, India
[email protected]

Abstract—Recently, deep learning methods are employed for [2]. So, MR image restoration by noise suppression is still an
image restoration tasks. An unsupervised learning technique is active research topic.
appropriate for many real time applications due to the scarcity Various MR image restoration techniques can be broadly
of a large amount of data for training. The conventional deep
image prior (DIP) is a CNN based denoiser prior that perform classified into traditional, and Deep Learning (DL) based
different image restoration tasks by using only a single degraded methods. Again, the traditional methods can be categorised
image. Alternating Direction Method of Multipliers (ADMM) into spatial domain or transform based filters. Many of the
framework over a standard sub-gradient method has already traditional filters use statistical techniques such as linear min-
been proposed with DIP method. Inspired by this, we propose imum mean square error (LMMSE) and maximum likelihood
a variant of ADMM-DIP method for enhancing single coil
magnitude magnetic resonance (MR) images. It is well known estimation (MLE) [2], [3]. Bilateral filter [4] is an edge
that the noise distribution in single coil magnitude MR images preserving non-iterative denoising method. The Non Local
is stationary Rician. We achieve the Rician noise removal from Means (NLM) filter [5] removes noise by utilizing self-spatial
single MR image by utilizing the combined effect of MSE, KL similarities in the image itself and leveraging redundancy of
divergence and perceptual loss functions. Also, the attention the adjacent pixels. Later, different variants of NLM like
guided dense upsampling network (AUNet) was engaged as
the CNN denoiser prior. Our experiments on simulated MR unbiased NLM (UNLM) [6], Adaptive NLM (ANLM) [7]
images indicate a better performance of the proposed method. were proposed for MR image denoising. BM3D [8] is a robust
We evaluated different denoising methods both qualitatively and algorithm with different parameters that can be fine-tuned to
quantitatively. achieve fine denoising effect.
Index Terms—Medical imaging, MR image, Denoising, CNN Recently, DL has emerged as the state-of-the-art (SoTA)
prior, ADMM, KL Divergence, Perceptual loss.
method for image restoration tasks. The different DL models
used for image restoration includes Convolutional Neural
I. I NTRODUCTION
Networks (CNNs) [9], Auto-encoders (AEs) [10], and Gener-
Image restoration is an inevitable pre-processing task in ative Adversarial Networks (GANs) [11]. DnCNN model uses
many image processing applications such as registration, supervised learning with CNN model for image restoration
segmentation, and object detection.Image restoration aims to in which residual modules and batch normalization improve
retrieve the clean image from the distorted observations. Also, the algorithm execution speed and performance [9]. Jiang et
the preservation of fine structures of the denoised image is al. [12] proposed a multi-channel DnCNN to denoise MR
vital in the medical field. [1]. images with known or unknown noise levels. Tripathi et al.
In modern medicine,Magnetic Resonance Imaging (MRI) [13] employed residual learning to denoise synthetic MRI. The
is a prominent non-intrusive imaging tool having no radiation network has been trained on Rician corrupted MR images at
risks involved. It is used for diagnosis and treatment of differ- higher noise levels and then used for blind MR denoising. AE
ent diseases, particularly in brain and spinal cord. However, with skip connections [10] was used to test Gaussian corrupted
noise generation during image acquisition is unavoidable in T1 weighted brain MRIs. Spatio-temporal denoising of brain
MRI. These noise patterns in MR images varies with different MRI was carried out by Benou et al. [14] using ensembles of
parameters such as number of coils used and signal strength deep neural networks.
The supervised learning-based methods require a large
∗ Corresponding Author(ORCID: 0000-0002-6738-7786) training dataset having clean images. For medical images, such
978-1-6654-2337-3/21/31.00 © 2021 IEEE a dataset creation is not feasible. Hence, the unsupervised DL

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
 methods provide an alternative and efficient approach that does the real and imaginary components of the complex signal are
not require a large dataset with ground truth (GT) images but corrupted by zero-mean uncorrelated Gaussian noise with the
still produces comparable results from a single noisy image same variance σ 2 , which is modeled by a Rician distribution
blindly. [22]. The Rice probability density function is given by:
Lehtinen et al. [15] trained Noise2Noise on individualistic
(R2 + V 2 )
   
R RV
pairs of noisy images and learned to predict a noisy image P (R|V, σ) = 2 exp − I0 (1)
from another noisy image. The trained model can predict the σ 2σ 2 σ2
expected value for the distribution, which is the denoised pixel where I0 is a modified Bessel function of the first kind with
value for many realistic noise models considered, which turned order zero and V is the underlying noiseless signal. At high
out to be an innovative approach in the unsupervised denoising SNR, Rician noise approaches a Gaussian distribution and at
domain. low SNR, it is a Rayleigh distribution.
In [16], the deep image prior (DIP) method has been
introduced, incorporating a CNN as a generator network with III. M ETHODOLOGY
randomly initialized weights and the degraded image alone as In the field of image processing, various methods com-
input. The CNN layers themselves will capture the majority monly employed for acquiring the restored image can be
of the image statistics without training over a large dataset. broadly categorized as regularised reconstruction-based and
However, it is prone to an overfitting problem without explicit deep learning-based techniques. In this paper, we employ a
regularization terms. In [17], an explicit total variation (TV) regularized reconstruction based technique that results in an
regularization term is added in the formulation of [16] to optimization problem of the form shown below:
achieve further improvement in the image restoration tasks. Find l∗ as,
Moreover, Pasquale Cascarano et al. proposed ADM M − 1
argmin ||Hl − k||22 + R(l)
DIPT V [18] for image restoration by using Alternating Di- l 2
rection Method of Multipliers (ADMM) based minimization where the first term is the fidelity term that model the
for solving the optimization problem to obtain good denoising noise component present in the image, while the second term
performance. denotes the regularization term which provides the prior infor-
A lot of work is conceived for deep image denoising, but the mation about the image, i.e., its nature, sparsity or regularity.
advancements in MR image restoration are limited. Inspired by
the shortfall of the existing methods and the need for a better- A. Model
unsupervised method for MR image denoising, we proposed The attention-guided dense upsampling network called
a new variant of the DIP method in this paper. AUNet [19] has been adopted into the proposed method to
Our contributions in this paper can be summarised as emphasize the high-level and low-level image details. The
follows: network consists of an asymmetrically structured encoder-
• The proposed method uses the ADMM DIP framework decoder architecture as shown in Figure 1 (a), with dissimilar
with the combined loss function for unsupervised single encoder-decoder blocks depicted in Figure 1 (b) and (d).
MR image denoising. The encoder block has residual connections to enable noise
• We updated the CNN prior in the ADMM DIP method detailing. Attention guided dense upsampling block is used
by AUNet [19] in the proposed method. in the decoder section to eliminate information loss caused
• The loss function employed here has been augmented by bilinear upsampling, efficiently fuse high- and low-level
with a weighted combination of perceptual loss [20] and features, and emphasise the rich-information channels. More-
Kullback-Leibler (KL) Divergence regularization term for over, channel-wise attention is employed in the Attention -
optimal denoising effect. Upsampling (AU) block to emphasize the detailing in the
• We compared the proposed method with different variants image to the region of interest (ROI) for improved perceptual
of DIP framework ( [16], [17], [21] and [18]) for MR quality. The final layer of the segmentation-based AUNet has
image enhancement. been changed to a 1 × 1 convolution layer to generate a single
The rest of the paper is organized as follows: Section II channel output image. Even though the network is similar to
provides the basics of noise characteristics in MRI. Section the UNet architecture involved in the original model [16], it
III details the methodology. Next, the results obtained with is computationally intensive due to the involvement of varied
the proposed method and its comparison with existing methods encoder and decoder block structures along with residual and
are presented in Section IV. Finally, Section V concludes the attention-based learning.
paper.
B. Loss Function
II. BACKGROUND 1) Perceptual Loss [23]: The VGG Perceptual Loss has
Usually, the noise patterns in natural images are considered been incorporated in this experimental setup to improve the
signal-independent Gaussian noise. In contrast, the noise in perceptual quality of the denoised images. This loss is derived
MRI is dependent on the image signal itself. The noise occurs from the ReLU activation layer of the pre-trained 19 layer
in the observed single-coil magnitude MR image R, where VGG network

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
 struction modelled with a probability distribution. With fewer
iterations, the KL-Divergence minimization allows for faster
convergence.
The ADMM [25] framework is utilized in the proposed
method as in [18] for optimization problem minimization. The
ADMM framework enables embedding any prior information
with its modular structure by changing the regularizer related
sub term. Here we have enhanced the performance of the
ADM M − DIPT V model by replacing the explicit prior,
total - variation with weighted divKL and Perceptual loss
combination. The loss weights are assigned through trial and
error considering the values generated by the loss components
during the analysis.
IV. E XPERIMENTAL R ESULT
This section entails the results of the proposed method
on the synthetic MR image corrupted with Rician noise at
different noise levels. The results were evaluated through
qualitative and quantitative analysis of the resultant images.
The results were compared to other SoTA denoising methods.
The comparison has been carried out with respect to the
base model of DIP [16], the DIPKL [21] method, DIPT V
method [26], ADM M − DIPT V method [18], LMMSE [3]
and BM3D [8]. The image quality metrics used for quantitative
evaluation in this paper are: peak signal to noise ratio (PSNR)
, and mean Structural Similarity Index (SSIM). The higher
the PSNR value, the better the performance of the denoising
method. Also, mean SSIM has a value in the range of [0, 1]
with ’1’ indicating maximum similarity to the GT.
A. Experimental Setup
The model was built using PyTorch platform and the exper-
Fig. 1: Pictorial representation of deep encoder-decoder model imentation was carried out using a workstation equipped with
[19] used in the proposed method. (a) The encoder-decoder Nvidia quadro RTX-8000 GPU.
model (b) Encoder block (c) Upsampling block and (d) De- The study has been carried out on the T1 weighted MR
coder block images of the brain obtained from simulated brain database
BrainWeb [27].
2) KL Divergence: The Kullback-Leibler divergence [24] In our experiments, we used the following parameters for
(divKL ) quantifies how one probability distribution differs each of the methods:
from another, and it is also known as ’Relative Entropy’ and • DIP [16] :The 32 channel input is randomly initialized,
is used to extrude the extra information required to calculate and the number of iterations is assigned 5000 with a
the value of s when it is modeled with an approximating learning rate of 0.01. The central part of the UNet
distribution q̃(s) rather than the true distribution, q(s). has 4-channel skip connections. Leaky ReLU activation
The discretized divKL of q = qij and q̃ = q̃ij can be used and the MSE loss are engaged. The channel distri-
as the similarity or disparity measure between the restored bution in upsampling and downsampling sections are
image b and the GT image a of dimensions e × f given by [8, 16, 32, 64, 128].
• DIPKL [21] : In contrast to DIP , Mish is used as
e−1 f −1
X X qij activation function and a combination of MSE and divKL
KL(qkq̃) = qij ln (2) is used as loss function. The number of iterations is set
i=0 j=0
q̃ij
as 2000 with λKL = 20. The upsampling and downsam-
In the proposed ADMM-DIP framework based method, we pling layers are assigned a uniform channel distribution,
have augmented the regularization term, KL divergence. KL- [128, 128, 128, 128, 128].
divergence is a popular metric for developing IR algorithms • DIPT V [17] : With DIP [16] configuration, TV reg-
since it plays a crucial role in compliance with the axiom ularization is used in the objective function. The TV
that reducing the KL-divergence is equivalent to improving regularization coefficient assigned is 10−7 . The number
the likelihood function that is considered suitable for recon- of iterations is assigned as 5000.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
 • BM 3D : Since the noise level being considered is less
than 40, parametric settings include the d-distance being
set to zero and the block size being used is eight for each.
• LM M SE [3] : This method enables to denoise the MR
image based on a Rician noise distribution assumption.
The window size considered is the default [7, 7] and the
noise estimation method engaged is ’mode2N’(local order
2 moment). (a) (b)
• ADM M − DIPT V [18] : This method utilizes the U-
Net architecture with up channels and down channels
involving 128 channels each. Skip connections of channel
size 8 are assigned at every level. During the iterative
optimization phase, denoised image corresponding to the
maximum PSNR value is selected. Loss function: M SE
+ standard isotropic total variation.
• P roposed : AUNet denoiser prior with ReLU activation (c) (d) (e)
function at each level for the residual blocks. The loss
function involved is a weighted combination of the MSE,
perceptual loss and DivKL . The same procedure as
in ADM M − DIPT V is involved while selecting the
optimally denoised image.
• ADM MDIP : Same experimental setup as proposed
method but with UNet as CNN prior.
(f) (g) (h)

Fig. 3: The residual image of the denoised image in Fig.2. (a)

Noisy (σ = 20) (b) DIP [16] (c) DIPKL [21] (d) DIPT V
[17] (e) BM3D filter (f) LMMSE [3] (g) ADM M − DIPT V
[18](h) Proposed

B. Qualitative Evaluation
(a) (b) (c)
Higher perceptual quality is attained by the proposed
method as compared to the other methods under consideration,
as can be inferred from Figure 2. This method is instrumental
in retaining the essential fine image details, particularly in
medical diagnosis. The proposed method can recover the tex-
ture pattern of the original image relatively. Figure 3 indicates
that the proposed method has a loss of edge details from the
original MR image but outperforms the other methods, that
(d) (e) (f)
direct to further study in this experimental setup.

C. Quantitative Evaluation
Tables I and II encompass the quantitative analysis for the
proposed method compared to other methods. The proposed
method outperformed its peers in PSNR and mean-SSIM
metrics, corresponding to the visual analytical inferences.
(g) (h) (i)
Compared with other DIP variants for MR denoising, the
Fig. 2: The denoised output obtained with different methods proposed method outperformed them in terms of PSNR. The
on a Rician noise corrupted Brainweb image. (a) GT (b) Noisy LMMSE estimation was evaluated to comparison as a SoTA
(σ = 20) (c) DIP [16] (d) DIPKL [21] (e) DIPT V [17] (f) MR denoising method and downgraded by the proposed
BM3D filter (g) LMMSE [3] (h) ADM M − DIPT V [18](i) method. The BM3D, a conventional, natural image denoiser,
Proposed was used for MR image denoising, and the proposed method
outperformed it in terms of quantitative metrics.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
 D. Discussion TABLE II: Performance comparison of different denoising
methods using mean SSIM obtained for different noise lev-
The activation function used in the model has been analysed els(SSIM value for noisy image provided in the first row).
to study the variation created in performance by commonly
used activation functions like ReLU, Leaky ReLU, PReLU, Methods σ = 10 σ = 20 σ = 30 σ = 40
and Mish. We used the aforementioned experimental setup Noisy 0.3947 0.2722 0.1956 0.1421
for conducting this study. Table III shows that the proposed DIP [16] 0.9827 0.9592 0.9271 0.8894
DIPKL [21] 0.9842 0.9573 0.9270 0.8847
method achieved enhancement concerning structural detail re- DIPT V [17] 0.9856 0.9601 0.9303 0.8927
tainment and loss reduction when the ReLU activation function BM3D 0.8978 0.7878 0.7104 0.6547
is engaged in the architecture. Hence, the proposed denoising LMMSE [3] 0.9485 0.8972 0.8583 0.8330
method has employed ReLU. ADM M − DIPT V [18] 0.9872 0.9454 0.9302 0.8921
Proposed 0.9698 0.9574 0.9346 0.8854
As aforementioned, we upgraded the CNN prior with
AUNet [19] in the ADMM - DIP model to obtain the proposed TABLE III: Performance of the proposed method with differ-
model. Table IV compares the performance of the ADMM- ent activation functions in Rician noise (σ = 20) corrupted
DIP framework with the base architecture, i.e., UNet, to the Brainweb data.
proposed method involving AUNet. Replacing the UNet with
Activation PSNR SSIM
skip connections, with the AUNet as the CNN prior resulted
in notable advancement in the performance. ReLU 30.8591 0.9574
Leaky ReLU 30.5885 0.9551
Table V discusses the performance of the proposed method PReLU 30.4117 0.9555
with varied combinations of loss functions considered in this Mish 30.6261 0.9564
study. The proposed method has the best performance with the
combinational loss involving M SE, DivKL and perceptual
V. C ONCLUSION
loss. The image profiling shown in Figure 4 agrees with
the quantitative assessment carried out and indicates that the In this paper, we proposed a new variant of ADMM based
resultant image of the proposed method is more similar to the DIP method tailored for MR image denoising. The method
GT as compared to the other methods. The intensity variations combines the goodness of attention mechanism, residual learn-
correspond to the central region of the GT as marked in Figure ing, dense upsampling along with bilinear upsampling and a
2. (a) are followed by the proposed method output (Figure 4 combinational loss function involving perceptual loss, MSE
(i)). The other DIP variants have a comparative performance loss and KL divergence with an ADMM based optimization
concerning the GT profile. model to achieve a superior image denoising performance
In DIP, the optimization to the denoised MR image for compared to its variants. We focused on stationary Rician
Rician noise level σ = 20 is obtained, at the 3411th iteration noise in a single coil magnitude MR image. Residual learning
in which each iteration takes approximately 0.2Sec, summing in the encoder section has enabled to focus on the noise
up to the optimal convergence at around 682 seconds. The content in the image. The experimentation has brought out
proposed method attains its maximum performance by 544 the advantages of utilizing the combinational loss function
iterations, but the time taken is 2709Sec which is a bit greater that enhances the PSNR as well as image detail retention. The
than the DIP model due to the perceptual loss function in proposed method has been computationally demanding due to
the objective function alongwith the computationally complex the involvement of perceptual loss and AUNet and thus tends
AUNet engaged as CNN prior. But this model goes beyond to be time consuming compared to other DIP variants as well
the performance attained by DIP and renders perceptually
enhanced MR images (Table II). The proposed model has TABLE IV: Comparison of the performance of the proposed
nearly 35 times the number of parameters as the DIP model. method with the ADMM -DIP method having UNet prior for
Rician noise σ = 20 in brainweb image.
TABLE I: Performance comparison of different denoising
methods using PSNR (dB) obtained for different noise lev- Method PSNR SSIM
els(PSNR value for noisy image provided in the first row). ADMM -DIP method 30.5961 0.9496
Proposed method 30.8591 0.9574
Methods σ = 10 σ = 20 σ = 30 σ = 40
TABLE V: Comparison of the performance of the proposed
Noisy 26.1441 20.1292 16.6268 14.0962 method with different loss function combinations for Rician
DIP [16] 27.9452 22.4714 19.499 17.7860
noise σ = 20 in brainweb image.
DIPKL [21] 27.8570 22.3988 19.6396 17.6590
DIPT V [17] 27.8809 22.4935 19.5137 17.8443
BM3D 31.5109 25.5314 21.9794 19.4795 Loss function PSNR SSIM
LMMSE [3] 33.3688 29.2201 27.1231 25.8636
LM SE 29.7407 0.9501
ADM M − DIPT V [18] 34.8628 29.0928 26.8543 23.5529
LM SE + LKL 30.6941 0.9590
Proposed 35.2113 30.8591 27.7746 25.8097
LM SE + LP erceptual 30.2640 0.9514
LM SE + LP erceptual + LKL 30.8591 0.9574

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
 [7] J. V. Manjón, P. Coupé, A. Buades, D. L. Collins, and M. Robles, “New
methods for mri denoising based on sparseness and self-similarity,”
Medical image analysis, vol. 16, no. 1, pp. 18–27, 2012.
[8] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by
sparse 3-d transform-domain collaborative filtering,” IEEE Transactions
on image processing, vol. 16, no. 8, pp. 2080–2095, 2007.
[9] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a gaussian
(a) (b) (c) denoiser: Residual learning of deep cnn for image denoising,” IEEE
transactions on image processing, vol. 26, no. 7, pp. 3142–3155, 2017.
[10] C. Bermudez, A. J. Plassard, L. T. Davis, A. T. Newton, S. M. Resnick,
and B. A. Landman, “Learning implicit brain mri manifolds with deep
learning,” in Medical Imaging 2018: Image Processing, vol. 10574.
International Society for Optics and Photonics, 2018, p. 105741L.
[11] Q. Yang, P. Yan, Y. Zhang, H. Yu, Y. Shi, X. Mou, M. K. Kalra,
Y. Zhang, L. Sun, and G. Wang, “Low-dose ct image denoising using a
generative adversarial network with wasserstein distance and perceptual
(d) (e) (f) loss,” IEEE transactions on medical imaging, vol. 37, no. 6, pp. 1348–
1357, 2018.
[12] D. Jiang, W. Dou, L. Vosters, X. Xu, Y. Sun, and T. Tan, “Denoising of
3d magnetic resonance images with multi-channel residual learning of
convolutional neural network,” Japanese journal of radiology, vol. 36,
no. 9, pp. 566–574, 2018.
[13] M. Tian and K. Song, “Boosting magnetic resonance image denoising
with generative adversarial networks,” IEEE Access, vol. 9, pp. 62 266–
62 275, 2021.
(g) (h) (i) [14] A. Benou, R. Veksler, A. Friedman, and T. R. Raviv, “Ensemble of
expert deep neural networks for spatio-temporal denoising of contrast-
Fig. 4: Profiling of the output images in Fig.1. (a) GT (b) Noisy enhanced mri sequences,” Medical image analysis, vol. 42, pp. 145–159,
2017.
(σ = 20) (c) DIP [16] (d) DIPKL [21] (e) DIPT V [17] (f) [15] J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala,
BM3D filter [28] (g) LMMSE [3] (h) ADM M − DIPT V and T. Aila, “Noise2noise: Learning image restoration without clean
[18](i) Proposed data,” arXiv preprint arXiv:1803.04189, 2018.
[16] D. Ulyanov, A. Vedaldi, and V. Lempitsky, “Deep image prior,” in
Proceedings of the IEEE conference on computer vision and pattern
recognition, 2018, pp. 9446–9454.
as the statistical denoising methods. Further improvement can [17] J. Liu, Y. Sun, X. Xu, and U. S. Kamilov, “Image restoration using total
be achieved by architectural and parameter tuning, which is a variation regularized deep image prior,” in ICASSP 2019 - 2019 IEEE
work under progress. International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2019, pp. 7715–7719.
C ONFLICT OF INTEREST [18] P. Cascarano, A. Sebastiani, and M. C. Comes, “Admm-diptv: combining
total variation and deep image prior for image restoration,” 2020.
None of the authors has any financial or personal relation- [19] H. Sun, C. Li, B. Liu, Z. Liu, M. Wang, H. Zheng, D. D. Feng,
and S. Wang, “Aunet: Attention-guided dense-upsampling networks for
ships that could inappropriately influence or bias the content breast mass segmentation in whole mammograms,” Physics in Medicine
of the paper. & Biology, vol. 65, no. 5, p. 055005, 2020.
[20] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta,
ACKNOWLEDGMENT A. Aitken, A. Tejani, J. Totz, Z. Wang et al., “Photo-realistic single
image super-resolution using a generative adversarial network,” in
This work was supported by the Science and Engineering Proceedings of the IEEE conference on computer vision and pattern
Research Board (Department of Science and Technology, recognition, 2017, pp. 4681–4690.
India) through project funding SRG/2019/001586. [21] R. Hari Kishan, C. Aneeta, and P. V. Sudeep, “Blind image restora-
tion with cnn denoiser prior,” in Proceedings of Second International
Conference on Data Science and Applications (ICDSA 2021), 2021.
R EFERENCES [22] H. Gudbjartsson and S. Patz, “The rician distribution of noisy mri data,”
[1] A. F. Sheta, “Restoration of medical images using genetic algorithms,” Magnetic resonance in medicine, vol. 34, no. 6, pp. 910–914, 1995.
in 2017 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), [23] J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time
2017, pp. 1–8. style transfer and super-resolution,” in European conference on computer
[2] P. V. Sudeep, P. Palanisamy, C. Kesavadas, and J. Rajan, “Nonlocal linear vision. Springer, 2016, pp. 694–711.
minimum mean square error methods for denoising mri,” Biomedical [24] J. M. Joyce, Kullback-Leibler Divergence. Berlin, Heidelberg: Springer
Signal Processing and Control, vol. 20, pp. 125–134, 2015. Berlin Heidelberg, 2011, pp. 720–722.
[3] S. Aja-Fernández, C. Alberola-López, and C.-F. Westin, “Noise and [25] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, 2011.
signal estimation in magnitude mri and rician distributed images: a [26] J. Liu, Y. Sun, X. Xu, and U. S. Kamilov, “Image restoration using total
lmmse approach,” IEEE transactions on image processing, vol. 17, no. 8, variation regularized deep image prior,” 2018.
pp. 1383–1398, 2008. [27] C. A. Cocosco, V. Kollokian, R. K.-S. Kwan, G. B. Pike, and A. C.
[4] C. Tomasi and R. Manduchi, “Bilateral filtering for gray and color Evans, “Brainweb: Online interface to a 3d mri simulated brain
images,” in Sixth international conference on computer vision (IEEE database,” in NeuroImage. Citeseer, 1997.
Cat. No. 98CH36271). IEEE, 1998, pp. 839–846. [28] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by
[5] A. Buades, B. Coll, and J.-M. Morel, “A non-local algorithm for image sparse 3-d transform-domain collaborative filtering,” IEEE Transactions
denoising,” in 2005 IEEE Computer Society Conference on Computer on Image Processing, vol. 16, no. 8, pp. 2080–2095, 2007.
Vision and Pattern Recognition (CVPR’05), vol. 2. IEEE, 2005, pp.
60–65.
[6] J. V. Manjón, J. Carbonell-Caballero, J. J. Lull, G. Garcı́a-Martı́,
L. Martı́-Bonmatı́, and M. Robles, “Mri denoising using non-local
means,” Medical image analysis, vol. 12, no. 4, pp. 514–523, 2008.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on November 04,2022 at 13:45:44 UTC from IEEE Xplore. Restrictions apply.
View publication stats