adaptive noise driven total variation filtering for magnitude mr image denosing

21
Adaptive Noise Driven Total Variation Filtering for Magnitude MR Image Denosing Nivitha Varghees. V, M. Sabarimalai Manikandan & Rolant Gini

Upload: indian-institute-of-technology-bhubaneswar

Post on 04-Jul-2015

393 views

Category:

Education


0 download

DESCRIPTION

Linear models (e.g. Gaussian filter) Nonlinear models (e.g. Median filter) DCT and SVD filters Neighborhood filters Non-Local filters Multiscale LMMSE Bilateral filters Wavelet transforms Total variation (TV) filter

TRANSCRIPT

Page 1: Adaptive noise driven total variation filtering for magnitude mr image denosing

Adaptive Noise Driven Total Variation Filtering for Magnitude MR Image Denosing

Nivitha Varghees. V, M. Sabarimalai Manikandan & Rolant Gini

Page 2: Adaptive noise driven total variation filtering for magnitude mr image denosing

Outline

Introduction

MR Image Denoising Methods

Proposed Total Variation Filtering

Rician Noise Estimation Scheme

Experimental Results

Conclusion

Page 3: Adaptive noise driven total variation filtering for magnitude mr image denosing

non-invasive medical test that widely used by physicians to diagnose and treat pathologic conditions

provide clear and detailed structures of internal organs and tissues of human body

MRI to examine the brain, spine, cardiac, abdomen, pelvis, breast, joints (e.g., knee, wrist, shoulder, and ankle), blood vessels and other body parts

MRI scanner uses powerful magnets and radio waves to create pictures of the body

The main source of noise in MRI images is the thermal noise in the patient’s body [Now99].

Magnetic Resonance (MR)Imaging

Page 4: Adaptive noise driven total variation filtering for magnitude mr image denosing

The raw MR data generated from a MRI machine are corrupted by zero-mean Gaussian distributed noise with equal variance

In MR reconstruction, the spatial-domain complex MR data are obtained from the frequency-domain raw MR data by taking an inverse Fourier transform (IFT)

For clinical diagnosis and analysis, the magnitude MR data is constructed by taking the square root of the sum of the square of the two real and imaginary data.

The construction of the magnitude MR data transforms the distribution of Gaussian noise into a Rician distributed noise

Rician Noise in Magnitude MR Images

Page 5: Adaptive noise driven total variation filtering for magnitude mr image denosing

Linear models (e.g. Gaussian filter)

Nonlinear models (e.g. Median filter)

DCT and SVD filters

Neighborhood filters

Non-Local filters

Multiscale LMMSE

Bilateral filters

Wavelet transforms

Total variation (TV) filter

Existing Image Denoising Methods

Page 6: Adaptive noise driven total variation filtering for magnitude mr image denosing

Gaussian filter performs well in the flat regions of images but do not well preserve the image edges.

Transform based methods fail to reproduce image details and often introduce artifacts

Neighborhood filters may distort fine edges and local geometries.

Non-local estimation method over smooth out image edges.

Wavelet based hard thresholding scheme introduce Gibbs oscillation near discontinuities.

Performance of Existing Methods

Page 7: Adaptive noise driven total variation filtering for magnitude mr image denosing

2E(u) = ( )

2u z R u

: norm of a vector,

z : observed noisy data,

u: desired unknown image to be restored,

λ: regularization parameter,

R (u) : regularization functionals

Formulation of Total Variation Filtering

Good edge preserving capability while simultaneously removing noise

Page 8: Adaptive noise driven total variation filtering for magnitude mr image denosing

Performance of the total variation filter for different values of regularization parameter

. (a) Original MRI image, (b) Image with Rician noise with σ= 10), (c) Restored image

with ʎ= 10, (d) Restored image with ʎ = 15, (e) Restored image with ʎ = 25, (f)

Restored image with ʎ= 40.

Performance of TV Filtering for Fixed Value of ʎ

Page 9: Adaptive noise driven total variation filtering for magnitude mr image denosing

Proposed Total Variation Filtering

Noisy Image Restored Image (ʎ)

Noisy Image Restored Image

ʎ

(a) Traditional TV Filtering Approach

(b) Propose Noise Level Driven TV Filtering Approach

Page 10: Adaptive noise driven total variation filtering for magnitude mr image denosing

where, Io denotes the 0th order modified Bessel function of the first kind,

σ2 is the noise variance in the MRI image,

A is the signal amplitude of the clean image,

m denotes the MR magnitude variable,

u (.) is the unit step Heaviside function.

Rician Distribution of Noisy MRI Data

The Rician probability density function (PDF) of the corrupted magnitude MR image intensity m is given as

2 2

02 2 2( , ) ( ),

2

m m A Amp m A e I u m

Page 11: Adaptive noise driven total variation filtering for magnitude mr image denosing

For magnitude MR images with large background, the Rician distribution is reduced to a Rayleigh distribution with the probability density function (PDF):

In the bright regions of the MR image where the SNR is assumed to be large, the Rician distribution can be approximated using a Gaussian PDF:

2

2 2( , ) ( )

2

m mp m A e u m

Noise Models for Rician Noise Estimation

Page 12: Adaptive noise driven total variation filtering for magnitude mr image denosing

Variance of noise in magnitude MR image is computed as

the regularization parameter of TV is adapted based on

the estimated noise standard deviation for a given image

Rician Noise Estimation Using Local Variances

Page 13: Adaptive noise driven total variation filtering for magnitude mr image denosing

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 360

2

4

6

8

10

12

14

16

18

20

22

24

26

28

30

32

34

36

noise standnard deviation, n

est

ima

ted

no

ise

sta

nd

ard

de

via

tion

,

Estimation of standard deviation of noise using the mode of the locally estimated variance. • corresponds to the average of the estimated standard deviations obtained for all test images for each noise standard deviation.

Performance of Rician Noise Estimator

Page 14: Adaptive noise driven total variation filtering for magnitude mr image denosing

Performance of Rician Noise Estimator

Page 15: Adaptive noise driven total variation filtering for magnitude mr image denosing

(a) Original MR image. (b) MR image with Rician noise with standard deviation σ= 20. (c) Restored image using proposed adaptive TV filtering approach. (d) Restored image using the multi-scale LMMSE approach. (e) Restored image using the non-local filtering approach. (f) Restored image using the bilateral filtering approach.

Performance of MRI Denoising Methods

Page 16: Adaptive noise driven total variation filtering for magnitude mr image denosing

σ=10 σ=15 σ=20 σ=25

MOS MSE SSIM MOS MSE SSIM MOS MSE SSIM MOS MSE SSIM

Noisy 4.3 130.3 0.506 2.3 303.8 0.37 1.3 551.7 0.29 1 874.4 0.24

NLM 4.8 79.6 0.639 4.3 180.1 0.55 3.7 326.9 0.49 3 522.7 0.45

Bilateral filter

4.7 89.98 0.617 3.7 205.4 0.51 2.8 369.1 0.49 2.3 579.4 0.38

Multiscale LMMSE

5 81.6 0.636 4.7 186.4 0.55 3.2 337.3 0.50 2.2 536.3 0.46

Proposed Method

5 80.4 0.639 4.6 180.4 0.55 4.2 324.6 0.49 3.4 513.3 0.44

Performance of MR Image Denoising Methods

SSIM: Structural Similarity Measure

MSE: Mean Squared Error

MOS: Mean Opinion Score

Page 17: Adaptive noise driven total variation filtering for magnitude mr image denosing

In this work, we studied the effect of regularization parameter on TV denoising

Here, the regularization parameter is adapted based on the noise level in magnitude MR images

The noise level is computed using local variances of image

The performance of different denoising algorithms such as NLM, bilateral, multiscale LMMSE approach, and proposed method is evaluated in terms subjective test and objective mertics

Experiments showed that the proposed method provides significantly better quality of denoised images as compared to that of the existing methods

Conclusion

Page 18: Adaptive noise driven total variation filtering for magnitude mr image denosing

R. M. Henkelman, “Measurement of signal intensities in the presence of noise in MR images,” Med. Phys., vol. 12, no. 2, pp. 232-233, 1985.

L. Kaufman, D. M. Kramer, L. E. Crooks, and D. A. Ortendahl, “Measuring signal-to-noise ratios in MR imaging,” Radiology, vol. 173, pp. 265-267, 1989.

G. Gerig, O. Kubler, R. Kikinis, and F. Jolesz. “Nonlinear anisotropic filtering of MRI data,” IEEE. Trans. Med. Img., vol. 11, no. 2, pp. 221- 232, 1992.

H. Gudbjartsson and S. Patz, ”The Rician distribution of noisy MRI data,”Magnetic Resonance in Medicine, vol. 34, no. 6, pp. 910-914, 1995.

A. Crdenas-Blanco, C. Tejos, P. Irarrazaval, and I. Cameron, “Noise in magnitude magnetic resonance images,” Concepts in Magnetic Resonance, vol. 32, no. 6, pp. 409-416, 2008.

S. Aja-Fernandez, C. Alberola-Lopez, and C.-F.Westin, “Noise and signal estimation in magnitude MRI and Rician distributed images: a LMMSE approach,” IEEE Trans. Image Process., vol. 17, no. 8, pp. 1383-1398, 2008.

References

Page 19: Adaptive noise driven total variation filtering for magnitude mr image denosing

A. Samsonov and C. Johnson. “Noise adaptive anisotropic diffusion filtering of MR images with spatially varying noise levels,” Magn. Reson. Med., vol. 52, pp. 798-806, 2004.

L. Rudin, S. Osher and E. Fatemi, “Nonlinear total variation based noise removal algorithms,” Physica D, vol. 60, pp. 259-268, 1992.

R. Acar and C. R. Vogel, “Analysis of total variation penalty methods for ill-posed problems,” Inverse Prob., vol. 10, 1217-1229, 1994.

A. Chambolle and P. L. Lions, “Image recovery via total variational minimization and related problems,” Numer. Math., vol. 76, pp. 167-188, 1997.

T. F. Chan, S. Osher, and J. Shen, “The digital TV filter and nonlinear denoising,” IEEE Trans. Image Process., vol. 10, no. 2, pp. 231-241, 2001.

A. Buades, B. Coll, and J.-M. Morel, “A review of image denoising algorithms, with a new one,” Multiscale Modeling and Simulatio, vol. 4, pp. 490-530, 2005.

A. Buades, B. Coll, and J.-M. Morel, “A non-local algorithm for image denoising,” IEEE Proc. Comp. Visi. Pat. Recog., vol. 2, pp. 60-65, 2005.

References

Page 20: Adaptive noise driven total variation filtering for magnitude mr image denosing

P. Perona, and J. Malik, “Scale-space and edge detection using anisotropic diffusion,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 12, no. 7, pp. 629-639, 1990.

J. Weickert, “Coherence-enhancing diffusion of colour images,” Image and Vision Computing, vol. 17, pp. 201-212, 1999.

L. Zhang et al., “Multiscle LMMSE-based image denoising with optimal wavelet selection, IEEE Trans. on Circuits and Systems for Video Technology, vol. 15, pp. 469-481, 2005.

A. Buades, B. Coll, and J. M. Morel “A non-local algorithm for image denoising,” IEEE Computer Vision and Pattern Recognition (CVPR), vol. 2, pp. 60-65, 2005.

Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600-612, 2004.

References

Page 21: Adaptive noise driven total variation filtering for magnitude mr image denosing