DWI Denoising Software

Download DWI Denoising Software

During last years, we proposed several denoising methods dedicated to 4D DWI. In order to make easier the use of these denoising filters, we decided to develop a Matlab-based package with GUI containing all the proposed methods. In addition, we included SPM routines for NIFTI reading and writing to propose an efficient format compatibility with standard software. Finally, we enable the processing of several files in 2-clicks to make easier the processing of large database. The different filters cover a large range of clinical situations and you should be able to find a way to denoise your 4D DWI images with this package.

Matlab package

The Matlab package DWIdenoisingPackage_r01 contains six denoising filters for 4D DWI compiled for Linux (64 bits), Windows (64 bits) and MAC (64bits) .

All the methods are copyrighted and are the property of Pierrick Coupé and Jose Manjon.

Installation instructions:

• 1. Download the package

  2. Extract the files from .zip

  3. Open Matlab

  4. Select the directory of the DWIdenoisingPackage_r01 as current directory

  5. Run the script MainDWIDenoising.m (by opening it and pressing ‘F5′ or by writing ‘MainDWIDenoising’ in the prompt).

  6. Follow the instructions of the GUI

     • Select one or several 4D NIFTI image(s) (you can try the image provided as example in the directory sampledata in order to verify your architecture compatibility)

• • • Press run

    • Wait and check the result on the figure. The denoising image(s) should be saved as .nii in the same directory than the noisy image(s) with the selected suffix (e.g., _denoised). One figure will be displayed by selected file.

Filters description and usage:

In this section, I tried to coarsely define the underlying assumptions of each method and the possible applications for fully-automatic processing. It might occur that your problem is non listed in the proposed applications. In such situation, the smoothing parameter can be tune to be adjusted for your application.

Recommended filters: LPCA and AONLM

LPCA filter:

Description: The most efficient denoising filter proposed in the package for DWI is the Local PCA [8]. This filter has been designed for spatially varying noise typically presents in parallel imaging. By using integrated noise estimation, this filter is fully automatic and quiet robust. It is the only proposed filter taking into account the 4D nature of DWI dataset.

Utilization: This filter is theoretically dedicated to DWI acquired with

• • multi-coils (non stationary noise)

  • SENSE reconstruction (SENSE results in Rican noise, GRAPPA results in non-central Chi noise)

  • without partial Fourier (non correlated noise)

In this case, the Rician option should be activated. This filter obtained good results in many clinical setups especially on SENSE without partial Fourier. In practice, LPCA will be more robust than other filter face of partial Fourier. LPCA filter can be used with Gaussian model. Finally, for GRAPPA reconstruction no adapted methods are proposed in this package. However, by using Gaussian noise model (i.e., non-activation of Rician option) good results can be obtained. Only the intensity bias correction will not be achieved.

AONLM filter:

Description: The second denoising filter proposed in the package is the Adaptive Optimized Nonlocal Means (AONLM) [4]. This filter has been designed for spatially varying noise typically presents in parallel imaging. By using integrated noise estimation, this filter is fully automatic and quiet robust. This filter was included in the method which won the ISBI 2013 HARDI and DTI challenges [9].

Utilization: This filter is theoretically dedicated to MRI acquired with

• • multi-coils (non stationary noise)

  • SENSE reconstruction (SENSE results in Rican noise, GRAPPA results in non-central Chi noise)

  • without partial Fourier (non correlated noise)

In this case, the Rician option should be activated. This filter obtained good results in many clinical setups especially on SENSE without partial Fourier. In practice, AONLM will be more robust than ONLM face of partial Fourier. AONLM filter can be used with Gaussian model. Finally, for GRAPPA reconstruction no adapted methods are proposed in this package. However, by using Gaussian noise model (i.e., non-activation of Rician option) good results can be obtained. Only the intensity bias correction will not be achieved.

Other filters: ONLM, ONLM multiresolution, ODCT, PRINLM

ONLM filter:

Description: This package includes the Optimized Nonlocal Means (ONLM) [1]. This filter is currently the state-of-the-art for 3D MRI denoising and has been well-validated. The impact of this filter on segmentation or cortical surface extraction has been investigated (see here for details). A Rician-adapted version (ORNLM) has been proposed in [3] to handle the intensity bias introduced by Rician noise. This filter requires the estimation of the noise variance to be fully automatic. In the package, we use our RMAD method [5] in order to estimate the noise variance. The ONLM filter is included in several software such as MedINRIA and Minctool.

Utilization: This filter is theoretically dedicated to MRI acquired with

• • 1 coil (stationary noise)

  • without partial Fourier (non correlated noise)

In this case, the Rician option should be activated. In case of zero-padding in k-space or partial Fourier acquisition, the noise variance case be under-estimated, we suggest to use increase smoothing parameter or to use LPCA or AONLM which may be more robust to correlated noise. In case of SENSE or GRAPPA acquisition, we suggest to use LPCA or AONLM.

Multi-resolution ONLM filter:

Description: This package includes the Multi-resolution Optimized Nonlocal Means (ONLM) [1]. Based on a similar approach than ONLM, this filter is more efficient on very low SNR image compared to ONLM. A Rician-adapted version has been proposed in order to handle the intensity bias introduced by Rician noise. This filter requires the estimation of the noise variance to be fully automatic. In the package, we use our RMAD method [5] in order to estimate the noise variance.

Utilization: This filter is theoretically dedicated to MRI acquired with

• • 1 coil (stationary noise)

  • without partial Fourier (non correlated noise).

In this case, the Rician option should be activated. In case of zero-padding in k-space or partial Fourier acquisition, the noise variance case be under-estimated, we suggest to use increase smoothing parameter or to use LPCA or AONLM which may be more robust to correlated noise. In case of SENSE or GRAPPA acquisition, we suggest to use LPCA or AONLM.

ODCT filter:

Description: This package also includes the Oracle-based DCT filter [6]. This filter obtained really good results and is the fastest proposed methods. A Rician-adapted version has been proposed in order to handle the intensity bias introduced by Rician noise. This filter requires the estimation of the noise variance to be fully automatic. In the package, we use our RMAD method [5] in order to estimate the noise variance.

Utilization: This filter is theoretically dedicated to MRI acquired with

• • 1 coil (stationary noise)

  • without partial Fourier (non correlated noise)

In this case, the Rician option should be activated. In case of zero-padding in k-space or partial Fourier acquisition, the noise variance case be under-estimated, we suggest to use increase smoothing parameter or to use LPCA or AONLM which may be more robust to correlated noise. In case of SENSE or GRAPPA acquisition, we suggest to use LPCA or AONLM.

PRINLM filter:

Description: This package also includes the Prefiltered Rotationally Invariant NonLocal Means filter [6]. This approach used ODCT results as preprocessing step. This filter obtained the best results over the proposed methods and is the second fastest. A Rician-adapted version has been proposed in order to handle the intensity bias introduced by Rician noise. This filter requires the estimation of the noise variance to be fully automatic. In the package, we use our RMAD method [5] in order to estimate the noise variance.

Utilization: This filter is theoretically dedicated to MRI acquired with

• • 1 coil (stationary noise)

  • without partial Fourier (non correlated noise)

In this case, the Rician option should be activated. In case of zero-padding in k-space or partial Fourier acquisition, the noise variance case be under-estimated, we suggest to use increase smoothing parameter or to use LPCA or AONLM which may be more robust to correlated noise. In case of SENSE or GRAPPA acquisition, we suggest to use LPCA or AONLM.

Acknowledgments:

This package benefits from the use of nifti routines from the well-known SPM software (http://www.fil.ion.ucl.ac.uk/spm/). All these routines remain the property of the Wellcome Trust Centre for Neuroimaging. In addition, this package benefits from the use of 3D wavelet Matlab toolbox of the Polytechnic University of Brooklyn (http://eeweb.poly.edu/iselesni/WaveletSoftware/index.html). This Matlab toolbox remains the property of authors.

References:

• 1. P. Coupé, P. Yger, S. Prima, P. Hellier, C. Kervrann, C. Barillot. An Optimized Blockwise NonLocal Means Denoising Filter for 3-D Magnetic Resonance Images. IEEE Transactions on Medical Imaging, 27(4):425–441, 2008.

  2. P. Coupé, P. Hellier, S. Prima, C. Kervrann, C. Barillot. 3D Wavelet Subbands Mixing for Image Denoising. International Journal of Biomedical Imaging, 2008.

  3. N. Wiest-Daesslé, S. Prima, P. Coupé, S.P. Morrissey, C. Barillot. Rician noise removal by non-Local Means filtering for low signal-to-noise ratio MRI: applications to DT-MRI. In MICCAI’08, pages 171–179, 2008.

  4. J. V. Manjon, P. Coupé, L. Martí-Bonmatí, D. L. Collins, M. Robles. Adaptive Non-Local Means Denoising of MR Images with Spatially Varying Noise Levels. Journal of Magnetic Resonance Imaging, 31(1):192–203, 2010.

  5. P. Coupé, J. V. Manjon, E. Gedamu, D. Arnold, M. Robles, D. L. Collins. Robust Rician Noise Estimation for MR Images. Medical Image Analysis, 14(4) : 483–493, 2010.

  6. J. V. Manjon, P. Coupé, A. Buades, D. L. Collins, M. Robles. New Methods for MRI Denoising based on Sparseness and Self-Similarity. Medical Image Analysis, 16(1): 18-27, 2012.

  7. P. Coupé, J. V Manjon, M. Robles, D. L. Collins. Adaptive multiresolution non-local means filter for three-dimensional magnetic resonance image denoising. IET Image Processing, 6(5): 558-568, 2012.

  8. J. V. Manjon, P. Coupé, L. Concha, A. Buades, D. L. Collins, M. Robles. Diffusion Weighted Image Denoising using overcomplete Local PCA. PLoS ONE 8(9): e73021. doi:10.1371/journal.pone.0073021.

  9. E. Garyfallidis, S. St-Jean, M. Paquette, P. Coupé, M. Descoteaux. Constrained spherical deconvolution on signal and ODF values. ISBI'13, HARDI Reconstruction Challenge, page 17, 2013.