Novel Denoising

8/3/2019 Novel Denoising

1/5

International Journal of Computer Science and CommunicationVol. 2, No. 2, July-December 2011, pp. 359-363

ABSTRACT

A NOVEL APPROACH NOISE FILTRATION FOR MRI IMAGE SAMPLE IN

MEDICAL IMAGE PROCESSING

Lalitha Y. S.1* and Mrityunjaya V. Latte2

1

Appa Institute of Engineering & Technology Gulbarga, Karnataka, India.2JSS Academy of Technical Education, Bangalore, India.

In this paper a modified spatial filtration approach is suggested for image denoising applications. The existing spatialfiltration techniques were improved for the ability to reconstruct noise-affected medical images. The developedmodified approach is developed to adaptively decide the masking center for a given MRI image. The conventionalfiltration techniques using mean, median and spatial median filters were analyzed for the improvement in modifiedapproach. The developed approach is compared with current image smoothening techniques. The proposed approachis observed to be more accurate in reconstruction over other conventional techniques.

Keywords: Spatial Filter, Image Denoising, Modified Spatial Filter, RMSE.

* Corresponding Author:[email protected]

1. INTRODUCTION

Digital image processing algorithms are generallysensitive to noise. The final result of an automatic visionsystem may change according whether the input MRIimage is corrupted by noise or not. Noise filters are thenvery important in the family of preprocessing tools.Inappropriate and coarse results may stronglydeteriorate the relevance and the robustness of acomputer vision application. The main challenge innoise removal consists in suppressing the corruptedinformation while preserving the integrity of fine

medical image structures. Several and well-establishedtechniques, such as median filtering are successfullyused in gray scale imaging.

Median filtering approach is particularly adaptedfor impulsive noise suppression. It has been shown thatmedian filters present the advantage to remove noisewithout blurring edges since they are nonlinearoperators of the class of rank filters and since theiroutput is one of the original gray values [1]-[2]. Theextension of the concept of median filtering to colorimages is not trivial. The main difficulty in defining arank filter in color image is that there is no naturaland unambiguous order in the data [3]-[4]. During the

last years, different methods were proposed to usemedian filters in color medical image processing [5]-[6].Whatever the vector filtering method, the challengeis to detect and replace noisy pixels whereas the relevantinformation is preserved. But it is recognized that insome MRI image areas most of vector filters blur thindetails and image edges [7]-[8]. Even if many works suchas Khriji and Gabbouj [9]. Generally impulse noisecontaminates medical images during data acquisition

by camera sensors and transmission in thcommunication channel. In [10] Chan and Nikolovproposed a two-phase algorithm. In the first phase othis algorithm, an adaptive median filter (AMF) is usedto classify corrupted and uncorrupted pixels; in thsecond phase, specialized regularization method iapplied to the noisy pixels to preserve the edges and noissuppression. The main drawback of this method is thathe processing time is very high because it uses a verylarge window size of 39X39 in both phases to obtain thoptimum output; in addition, more Complex circuitry ineeded for their implementation. In [11] Srinivasan andEbenezer proposed a sorting based algorithm in whicthe corrupted pixels are replaced by either the medianpixel or neighborhood pixel in contrast to AMF and otheexisting algorithms that use only median values foreplacement of corrupted pixels. At higher noise densitiethis algorithm does not preserve edge and fine detailsatisfactorily. In this paper a novel robust estimationbased filter is proposed to remove fixed value impulsnoise effectively. The proposed filter removes low to higdensity fixed value impulse noise with edge and detaipreservation upto a noise density of 90%. Recentlynonlinear estimation techniques are gaining popularity

for the problem of image denoising. The well-knownWiener filter for minimum mean-square error (MMSEestimation is designed under the assumption of widesense stationary signal and noise (a random process isaid to be stationary when its statistical characteristicare spatially invariant) [12]. For most of the natural MRimages, the stationary condition is not satisfied. In thpast, many of the noise removing filters were designedwith the stationary assumption. These filters removnoise but tend to blur edges and fine details. Thialgorithm fails to remove impulse noise in high frequenc
mailto:[email protected]:[email protected]


2/5

International Journal of Computer Science and Communication (IJCSC)360

regions such as edges in the MRI image. To overcomethe above mentioned difficulties a nonlinear estimationtechnique for the problem of medical image denoisinghas been developed based on robust statistics. Robuststatistics addresses the problem of estimation when theidealized assumptions about a system are occasionallyviolated. The contaminating noise in an image is

considered as a violation of the assumption of spatialcoherence of the medical image intensities and is treatedas an outlier random variable [12]. In [13] Kashyap andEom developed a robust parameter estimation algorithmfor the medical image model that contains a mixture ofGaussian and impulsive noise. In [12] a robust estimationbased filter is proposed to remove low to mediumdensity Gaussian noise with detail preservation. Thoughthee techniques were developed for filtration of Gaussianor impulsive noise they are been developed for gray levelimages and are not suitable for color images. In this papera modified approach to spatial median filter is proposedfor the noise removal in digital medical images. The

paper is further presented in six sections. Whereconventional spatial filtration methods and theirlimitations were presented in Section 2. Section 3 outlinesthe proposed modified median filtration approach forMRI images. The simulation observations were presentedin section 4.

2. SPATIAL FILTRATION

The simplest of smoothing algorithms is the Mean Filter.Mean filtering is a simple, intuitive and easy toimplement method of smoothing medical images, i.e.reducing the amount of intensity variation between onepixel and the next. It is often used to reduce noise inMRI images. The idea of mean filtering is simply toreplace each pixel value in an image with the mean(average) value of its neighbors, including itself. Thishas the effect of eliminating pixel values, which areunrepresentative of their surroundings. The mean valueis defined by,

1

1( )

N

i ii

MnF x xN =

= (1)

Where, N : number of pixels

xi : corresponding pixel value,

I : 1.. N.The mean filtration technique is observed to be lower

in maintaining edges within the images. To improve thislimitation a median filtration technique is developed.The median filter is a non-linear digital filteringtechnique, often used to remove noise from medicalimages or other signals. Median filtering is a commonstep in image processing. It is particularly useful toreduce speckle noise and salt and pepper noise. Its edge-preserving nature makes it useful in cases where edge

blurring is undesirable. The idea is to calculate thmedian of neighbouring pixels values. This can be donby repeating these steps for each pixel in the medicaimage.

a) Store the neighbouring pixels in an array. Thneighbouring pixels can be chosen by any kind oshape, for example a box or a cross. The array i

called the window, and it should be odd sized.b) Sort the window in numerical order

c) Pick the median from the window as the pixelvalue.

The median is defined by,

( )2

( )i i MdF x Median x= (2

where, i = 1. N. These filtration techniques were foundto be effective in gray scale images. When processed ovecolor images these filtration techniques give lesseperformance. To achieve accurate reconstruction o

medical image the median filtration technique imodified to spatial median filtration. The Spatial MediaFilter is a uniform smoothing algorithm with the purposof removing noise and fine points of medical image datwhile maintaining edges around larger shapes. The basialgorithm for spatial median filter is as outlined belowthe algorithm determines the Spatial Median of a set opoints, x1, ...,xN:

1. For each vector x, compute S, which is a set of thsum of the spatial depths from x to every othevector.

2. Find the maximum spatial depth of this set, Smax

3. Smax is the Spatial Median of the set of points.The spatial depth between a point and set of point

is defined by,

1

( )1( , ) 1

( 1)

Ni

deepth ii i

X xS X x

N X x== (3

The Spatial Median Filter is an unbiased smoothinalgorithm and will replace every point that is not thmaximum spatial depth among its set of mask neighborsTo eliminate this limitation in this paper a ModifiedSpatial Median Filter for MRI image denoising iproposed.

3. MODIFIED SPATIAL FILTER

In a Spatial Median Filter the vectors are ranked by somcriteria and the top ranking point is used to the replacthe center point. No consideration is made to determinif that center point is original data or not. The unfortunatdrawback to using these filters is the smoothing thaoccurs uniformly across the image. Across areas wherthere is no noise, original medical image data is removedunnecessarily. In the Modified Spatial Median Filter


3/5

A Novel Approach Noise Filtration for MRI Image Sample in Medical Image Processing 361

after the spatial depths between each point within themask are computed, an attempt is made to use thisinformation to first decide if the masks center point isan uncorrupted point. If the determination is made thata point is not corrupted, then the point will not bechanged. The proposed modified filtration works asexplained below:

1) Calculate the spatial depth of every point withinthe mask selected.

2) Sort these spatial depths in descending order.

3) The point with the largest spatial depth representsthe Spatial Median of the set. In cases where noiseis determined to exist, this representative point isused to replace the point currently located underthe center of the mask.

4) The point with the smallest spatial depth will beconsidered the least similar point of the set.

5) By ranking these spatial depths in the set indescending order, a spatial order statistic of depth

levels is created.

6) The largest depth measures, which represent thecollection of uncorrupted points, are pushed to thefront of the ordered set.

7) The smallest depth measures, representing pointswith the largest spatial difference among others inthe mask and possibly the most corrupted points,and they are pushed to the end of the list.

This prevents the smoothing by looking for theposition of the center point in the spatial order statisticlist. For a given parameter (where 1 mask size),which represents the estimated number of original points

under a mask of points. As stated earlier, points withhigh spatial depths are at the beginning of the list. Pixelswith low spatial depths appear at the end.

If center point c then current pixel MSF (, xi)= rc

elsif center point c > then current pixel MSF(, xi) = r1

else if c = 1 then, pixel canot be modified.

If the position of the center mask point appearswithin the first bins of the spatial order statistic list,then the center point is not the best representative pointof the mask, and it is still original data and should notbe replaced.

Two things should be noted about the use of inthis approach. When is 1, this is the equivalent to theunmodified Spatial Median Filter. When is equal tothe size of the mask, the center point will always fallwithin the first bins of the spatial order statistic andevery point is determined to be original. This is theequivalent of performing no filtering at all, since all ofthe points are left unchanged. The algorithm to detectthe least noisy point depends on a number of conditions.

First, the uncorrupted points should outnumber, or bmore similar, to the corrupted points. If two or morsimilar corrupted points happen in close proximity, thethe algorithm will interpret the occurrence as originadata and maintain the corrupted portions. While is aestimation of the average number of uncorrupted pointunder a mask of points, the experimental testing mad

no attempt to measure the impulse noise composition oan medical image prior to executing the filter.

4. RESULTS AND DISCUSSION

Different stages of filtered MRI images are shown inFig.1. To test the accuracy of the modified spatial mediafilter, a medical image with corruption applied bysome means is applied. To estimate the quality of reconstructed MRI image, first calculate the Root-MeanSquared Error between the original and the reconstructeimage. The Root-Mean-Squared Error (RMSE) for aoriginal image I and reconstructed MRI image R idefined by,

2

0 0

1( , ) ( , ) ( , )

Iw Ih

i jw h

RMSE I R I i j R i jI I = =

=

(4

The algorithm for the Modified Spatial Median Filte(MSMF) requires two parameters. The first parameteconsidered is the size of the mask to use for each filterinoperation. The second parameter, threshold , representthe estimated number of original points for any givensample under a mask. A collection of ten MRI images ovarious sizes was used in these tests. These images are variety of textures and subject matter. The texture othese MRI images impact on the threshold chosen than

the window mask size. The tests to determine the besmask size were conducted in this manner: 1. Each of thten MRI images in the collection was artificially distortewith =0.0, =0.05, =0.10, and =0.20 noise compositionresulting in 40 images 2. Each of the forty medical noisimages was then reconstructed using the SMF with massizes of N=3, N=5, and N=7 (the second argumentthreshold , is set to 1), resulting in 120 reconstructedmedical images. 3. The Root-Mean-Squared Error wacomputed between all 120 reconstructed MRI images andthe originals. The RMSE is a simple estimation score othe difference between two MRI images. An ideal RMSEwould be zero, which means that the algorithm correctly

identified each noisy point and also correctly derivedthe original data at that location in the signal. For thevaluation of the work a performance evaluation icarried out for various samples and the results obtainedwere as shown below. As seen in the figure 2, a massize of 3 clearly outperformed the other tested sizes o5 and 7. Neither the amount of noise, the size of the MRimage, nor the subject matter of the image effects onwhich mask size performed the best. Less thorough testwere run on higher mask sizes such as 9 and 11. Wit


4/5

International Journal of Computer Science and Communication (IJCSC)362

and sandy beaches. When comparing the ModifieSpatial Median Filter, the Spatial Median Filter, thMedian Filter, and the Mean Filter, a 2000 image subseof the 59,895 images were used.

The above figure 3. gives the comparison of all thfour-filtration techniques used in this approach. Thworst performing filter of the four tests was the Meanfilter. For all noise models containing at least = 0.1noise, it produced the least accurate results. Thunmodified Median Filter was only marginally bettethan the Mean filter. The most stable of the filters wathe Median Filter, which is the best accuracy across aleleven tested noise models. For noise models containin

= 0.15, we see from Fig. 3 that Modified Spatial MediaFilter produced the most accurate images.

5. CONCLUSIONS

This paper introduces two new filters for removinimpulse noise from images and shown how theycompare to other well-known techniques for noisremoval. First, common noise filtering algorithms werdiscussed. Next, a Spatial Median Filter was proposedbased on a combination of work on the Median Filteand the Spatial Median quantile order statistic. Seeingthat the order statistic could be utilized in order to mak

a judgment as to whether a point in the signal iconsidered noise or not, a Modified Spatial MedianStatistic is proposed. The Modified Spatial Median Filterequires two parameters: A window size and a thresholdT of the estimated non-noisy pixels under a mask. In thresults, the best threshold T to use in the Modified SpatiaMedian Filter and determined that the best threshold i4 when using a 33 window m ask size. Using these aparameters, this filter was included in a comparison othe Mean, Median, and Spatial Median Noise Filters. Inthe broad comparison of noise removal filters, it wa

Figure 3: Effectiveness of Noise Filters for

Various Noise Compositions

Figure 1: MRI Image

Spatial Median FilterMRI image

Modified Spatial MedianFilter MRI image

Original Image Noisy MRI image

Filtered MRI image Filtered Median

Figure 2: Comparisons of Mask Sizes and Average RMSE with

Various Noise Compositions

each increase in mask size, the RMSE of each testincreased. Most of the medical images are images ofvarious scenes, such as portrait shots, nature shots,animal shots, scenic shots of snow-capped mountains


5/5

A Novel Approach Noise Filtration for MRI Image Sample in Medical Image Processing 363

concluded that for images containing p = 0.15 noisecomposition, the Modified Spatial Median Filterperformed the best and that the Component MedianFilter performed the best over all noise models tested.

REFERENCE

[1] R. Hodgson, D. Bailey, M. Naylor, A. Ng, and S. McCNeil,

Properties, Implementations and Applications of Rankfilters, Image Vision Comput., 3, pp. 314, 1985.

[2] M. Cree, Observations on Adaptive Vector Filters for

Noise Reduction in Color Images, IEEE Signal Processing

Letters, 11, No. 2, pp. 140143, 2004.

[3] P. Lambert and L. Macaire, Filtering and Segmentation:

The Specificity of Color Images, Proc. Conference on Color

in Graphics and Image Processing , Saint-Etienne.

[4] I. Pitas and P. Tsakalides, Multivariate Ordering in

Color Image Filtering, IEEE Transactions on Circuits and

Systems for Video Technology, 1, pp. 247259, 1991.

[5] R. Lukac, B. Smolka, K. Plataniotis, and A.

Venetsanopoulos, Selection Weighted Vector DirectionnalFilters, Computer Vision and Image Understanding, 94, pp.

140167, 2004.

[6] M. Vardavoulia, I. Andreadis, and P. Tsalides, A New

Median Filter for Colour Image Processing, Pattern

Recognition Letters, 22, pp. 675689, 2001.

[7] A. Koschan and M. Abidi, A Comparison of Media

Filter Techniques for Noise Removal in Color Images

Proc. 7th German Workshop on Color Image Processing, 34

No. 15, pp. 6979, 2001.

[8] R. Lukac, Adaptive Vector Median Filtering, Patter

Recognition Letters, 24, pp. 18891899, 2002.

[9] L. Khriji and M. Gabbouj, Vector Medianrationa

Hybrid Filters for Multichannel Image Processing, IEE

Signal Processing Letters, 6, No. 7, pp. 186190, 1999.

[10] Raymond H. Chan, Chung-Wa Ho, and Mila Nikolova

Salt-and-Pepper Noise Removal by Median-Type Nois

Detectors and Detail-Preserving Regularization, IEE

Trans. Image Processing, 14, No. 10, 2005.

[11] K. S. Srinivasan and D. Ebenezer, New Fast an

Efficient Decision-Based Algorithm for Removal o

High-Density Impulse Noises, IEEE Signal Processin

Letters, 14, No. 3, 2007.

[12] Rabie, Robust Estimation Approach for Blin

Denoising, IEEE Trans. Image Processing, 14, No.11, pp

1755-1765, 2005.

[13] R. Kashyap and K. Eom, Robust Image Modelin

Techniques with an Image Restoration Application

IEEE Trans. Acoust. Speech Signal Process , 36, No. 8

pp.1313-325, 1988.

Documents

Novel Denoising