28-04-2014, 12:50 PM
Noise reduction in magnetic resonance images using adaptive non-local means filtering
[attachment=62604]
ABSTRACT
Proposed is a noise reduction method for magnetic resonance (MR)
images. This method can be considered a new adaptive non-local
means filtering technique since different weights based on the edgeness
of an image are applied. Unlike conventional noise reduction methods,
which typically fail in preserving detailed information, the proposed
method preserves fine structures while significantly reducing noise in
MR images. For comparing the proposed method with other noise
reduction methods, both a simulated ground truth data set and real
MR images were used. The experiment shows that the proposed
method outperforms conventional methods in terms of both restoration
accuracy and quality.
Introduction:
Magnetic resonance imaging (MRI) is a diagnostic
imaging technique that produces highly detailed information of the
interior of the human body. Usually, MR images suffer from noise
because of short scan times, weak signal strength, the T1/T2 effect and
main/RF field inhomogeneity. To remove this noise, Gaussian smooth-
ing, a weighted average of neighbouring pixels, can typically be
applied. However, this method tends to blur structures at high contrast
regions containing important information, such as vessels and the bound-
aries between different organs. To preserve these fine structures while
reducing noise, many methods such as total variation, anisotropic diffu-
sion and bilateral filtering have been developed [1–3]. These approaches
take an average of neighbouring pixels using weights depending on the
detailed information. These approaches, however, do not sufficiently pre-
serve the fine structures, especially in MR images. Recently, non-local
means (NLM) filtering has been introduced for reducing noise in MR
images [4]. However, the original NLM filtering method also tends to
blur detailed information if the weight is not appropriately determined.
In this Letter, we propose an adaptive NLM filtering technique that
preserves fine structures while reducing noise in MR images.
Experimental results:
We compared the proposed method with conven-
tional methods (Gaussian averaging, total variation [1], anisotropic dif-
fusion [2], bilateral filtering [3] and NLM [5]) both quantitatively and
qualitatively. For the quantitative evaluation, we used the simulated
brain database from BrainWeb [7], which includes noise-free images
and images with 9% noise. The parameters used for the experiments
were as follows. For σ, we used the optimal values at the relevant
noise levels with a 5 × 5 patch and 11 × 11 search window, as suggested
in [4]. For the edge detectors, we used the Sobel operator and Δn was set
to − 1, 0 or 1.
Conclusion:
In this study, we have demonstrated the effectiveness of
our proposed adaptive NLM filtering for MR images by overcoming
the limitations of conventional NLM filtering that may result in the
loss of fine structural information. We first estimated the edgeness of
an image and then adaptively applied a non-local means filter based
on this edgeness. Experiments show that our proposed method preserves
the important structures in MR images, such as vessels, while signifi-
cantly reducing noise.
[attachment=62604]
ABSTRACT
Proposed is a noise reduction method for magnetic resonance (MR)
images. This method can be considered a new adaptive non-local
means filtering technique since different weights based on the edgeness
of an image are applied. Unlike conventional noise reduction methods,
which typically fail in preserving detailed information, the proposed
method preserves fine structures while significantly reducing noise in
MR images. For comparing the proposed method with other noise
reduction methods, both a simulated ground truth data set and real
MR images were used. The experiment shows that the proposed
method outperforms conventional methods in terms of both restoration
accuracy and quality.
Introduction:
Magnetic resonance imaging (MRI) is a diagnostic
imaging technique that produces highly detailed information of the
interior of the human body. Usually, MR images suffer from noise
because of short scan times, weak signal strength, the T1/T2 effect and
main/RF field inhomogeneity. To remove this noise, Gaussian smooth-
ing, a weighted average of neighbouring pixels, can typically be
applied. However, this method tends to blur structures at high contrast
regions containing important information, such as vessels and the bound-
aries between different organs. To preserve these fine structures while
reducing noise, many methods such as total variation, anisotropic diffu-
sion and bilateral filtering have been developed [1–3]. These approaches
take an average of neighbouring pixels using weights depending on the
detailed information. These approaches, however, do not sufficiently pre-
serve the fine structures, especially in MR images. Recently, non-local
means (NLM) filtering has been introduced for reducing noise in MR
images [4]. However, the original NLM filtering method also tends to
blur detailed information if the weight is not appropriately determined.
In this Letter, we propose an adaptive NLM filtering technique that
preserves fine structures while reducing noise in MR images.
Experimental results:
We compared the proposed method with conven-
tional methods (Gaussian averaging, total variation [1], anisotropic dif-
fusion [2], bilateral filtering [3] and NLM [5]) both quantitatively and
qualitatively. For the quantitative evaluation, we used the simulated
brain database from BrainWeb [7], which includes noise-free images
and images with 9% noise. The parameters used for the experiments
were as follows. For σ, we used the optimal values at the relevant
noise levels with a 5 × 5 patch and 11 × 11 search window, as suggested
in [4]. For the edge detectors, we used the Sobel operator and Δn was set
to − 1, 0 or 1.
Conclusion:
In this study, we have demonstrated the effectiveness of
our proposed adaptive NLM filtering for MR images by overcoming
the limitations of conventional NLM filtering that may result in the
loss of fine structural information. We first estimated the edgeness of
an image and then adaptively applied a non-local means filter based
on this edgeness. Experiments show that our proposed method preserves
the important structures in MR images, such as vessels, while signifi-
cantly reducing noise.