patch-based nonlocal denoising for mri and ultrasound images
DESCRIPTION
Patch-based Nonlocal Denoising for MRI and Ultrasound Images. Xin Li Lane Dept. of CSEE West Virginia University. Outline. I come to see and be seen Motivation: nonlocal ( symmetry -related) dependency in medical images Technical Approach - PowerPoint PPT PresentationTRANSCRIPT
Patch-based Nonlocal Denoising for MRI and Ultrasound Images
Xin Li
Lane Dept. of CSEE
West Virginia University
Outline• I come to see and be seen• Motivation: nonlocal (symmetry-related) dependency in
medical images• Technical Approach
– Patch-based image modeling and geometric resampling – From locally linear embedding (LLE) to locally linear transform
(LLT) – Nonlocal denoising algorithm
• Experimental results– Synthetic images, Gaussian noise– MRI images, Rician noise– Ultrasound images, speckle noise
Big Picture: Computational ImagingQuality
Cost
Physical
Examples: SMASH/SENSE for fast MRISuper-resolution in PET imaging
High-dynamic-range (HDR) imaging
Computational
Motivation: Modeling Human-related Prior
Bilateral symmetry Shape boundary regularity
Patch-based Image Modeling • To overcome the curse of
dimensionality, we have to work at the middle ground between pixel-level and image-level
• An old concept with renewing interest– Vector quantization is
patch-based, JPEG used 8-by-8 patches (SP community)
– Patch-based recognition (CV community)
– Nonlinear dimensionality reduction (ML community)
P
P
Nonparametric: patch-basedvs.
Parametric: wavelet-based
Nonlocal Dependency
reflective symmetry translational symmetry
Beyond the reach of any localized models (MRF, wavelet-based, PDE-based)
Redundant Representation by Geometric Resampling
fliplr(x)x flipud(x) flipud(fliplr(x))
Collection of P-by-P patches
Exploiting Manifold Constraint
B4
B2
B3
B0
B1
RPP
k
tttw
10 BB
Nonlinear Dimensionality ReductionBy Locally Linear Embedding (LLE) Roweis and Saul, Science’2000
0WD
},...,,1{ 1 kwwdiag W],...,,[ 10 kBBBD
Sparsifying transform
t
Artificial third dimension t records the location information
Nonlocal Sparse Representation (NSR)
0FD
Approximated solution(3D FFT/DCT)
0WD
Optimal sparsifyingtransform (KLT)
B0 BkB1 Pack into3D Array D 3D-FFT
Thresholding
…
3D-IFFTPack into
3D Array D
B0 BkB1 …^ ^ ^
NSR Image Denoising Algorithm
Experimental Results on NSR• Computer-generated toy images, additive
White Gaussian noise– Illustrate the algorithm procedure and verify the
benefit of resampling
• MRI images, Rician noise– Benchmark: PDE-based scheme (total-variation
denoising)
• Ultrasound images, speckle noise– Benchmark: local schemes (SRAD, SBF, PDE)
Denoising Procedure Illustration by Toy Example
Noisy image Search similar patchesNoisy 3D array
LLT Thresholding
denoised 3D arrayDenoised image denoised patches
Benefit of Resampling
Translation only
Translation and 1 reflection
Translation and 2 reflections
Translation and 3 reflections
original noisy
NSR (ISNR=17.5dB)GSM (ISNR=13.3dB)
GSM: Gaussian Scalar Mixture in Wavelet space (state-of-the-art denoising scheme)
MRI Image Denoising
original Noisy (Rician, =30)
PDE scheme NSR scheme
Ultrasound Despeckling
Field-IISimulation
SBF(local)
11.2ˆ Q
40.2ˆ Q02.2ˆ Q
NSR (nonlocal) 2
ˆˆˆ NSRSBF xxx
Q̂ Ultrasound Despeckling Assessment Index (USDSAI)**Tay, P.C.; Acton, S.T.; Hossack, J.A., “A stochastic approach to ultrasound Despeckling,”ISBI’2006
Other (Non-medical) Applications of Nonlocal Sparse Representation
original Randomly-sampled(20% data)
RUPScheme*
griddatascheme
EM+NSRscheme
*Candes, E.J.; Romberg, J.; Tao, T., “Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency Information,” IEEE Trans. on Infor. Theory, pp. 489- 509, Feb. 2006
Concluding Remarks
• Symmetry – an important piece of prior information about human subjects
• Patch-based models enable us to better distinguish signal (pattern of interest) from noise using the tool of nonlocal sparsity
• Our experiments have shown the effectiveness of such models in a variety of imaging modalities and noise conditions
• Interest in NIH RFP: Innovations in Biomedical Computational Science and Technology (R01)