Breast cancer becoming the most occurring forms ofcancer in women, especially in developing countries as well as in developedcountries. Mammogram is accepted as best radiographic method to clinicalexamination in the early stages of illnesses for breast cancer. Diagnosticfeatures such as cysts and masses may be small or ill-defined, so mammogramimage quality is extremely important when used for interpretation anddiagnosis. Pre-processing is an importantpart of any imaging modalities to bring the image to that quality where it issuitable for analysis and extraction of significant data. Thispaper talks about, six denoising techniques which have greatsignificance in mammographic image analysis due to poor quality of mammograms andwith the help of mini-MIAS database, results were evaluated using many image metrics. Canceris abnormal change, in the genes which are responsible for regulating thegrowth of cells and keeping them healthy, where genes in each cell’s nucleus,which acts as the “control room” of each cell. Normally, the process of cellgrowth like old cells in our bodies replaces themselves through a healthy new.
Sometimeschanged cell gains the ability to keep dividing without control or order, producingmore and forming a tumor 1. Breast cancer is one of the most occurring formsof cancer in women, especially in India. Breast Cancer detection at initialstage is the only solution, to reduce the death rate among the women’s, soregular screening of mammograms for detection of masses is essential, so wide range oftreatment options can be available at early stages like less-aggressive surgery and adjuvanttherapy. 2Mammography is a medical imaging which uses a low-dose x-raysystem to analyze inside the breasts. A mammography examination is knownas mammogram, whichaids in the early detection and diagnosis of breast diseases in women.
Mammogram is accepted as an important adjunct to clinicalexamination in the care of patients with many common illnesses for breastcancer. It providesthe complete information about shape, structure and pathologies of breast whichhelps out in diagnosing breast cancer. Mostly mammograms are found with five abnormalities whichrequire further examination to classify it as benign or malign. Themammograms are confirmed as abnormal after the clinical diagnosis mammography whichis divided into two classes: benign (noncancerous) and malignant (cancerous). The abnormalities are Asymmetricbreast tissue, Mass, Micro calcification, Asymmetric density and Architecturaldistortion.
Except for mass and micro calcification all the other abnormalitiesare found in very rare cases known as invasive breast cancer.3 Diagnostic features such as a benign ormalignant tumor, or non-masses are exclusively normal tissues which may be small or ill-defined, somammogram image quality is extremely important when used for interpretation anddiagnosis.4In digital imaging, systems introduce various noises and artifacts whichare to be removed. The noise likelow contrast of the small tumors to the background can be hardly detected inmammography.
For this an image preprocessing techniques are adopted to reducethe noise level of the image, preserving the mammography structures andfeatures. Also denoising is more significant in image processing, analysis and applications;by reserving the details of an image and removing the random noise as far aspossible is the goal of image denoising approaches. In this work, a comparative study of severaldenoising techniques for mammogram is presented. The filters consideredare: 1) median filter, 2) adaptivemedian filter, 3) frost filter, 4) wavelet filter, 5) histogram equalization,6) Contrast Limited Adaptive Histogram Equalization.A traditional way toremove noise from image data is to employ spatial filters.
In Spatialfiltering the processed value for the current pixel depends on both itself andsurrounding pixels. Spatial filter are classified into non-linear and linear filters.Linearfilters process time-varying input signals to produce output signals, subjectto constraint of linearity whereas non-linear filters produces output signals, subjectto non-linearity especially in removal of certain types of noise that are notadditive.Median filtering is a nonlinear process which is useful inpreserving edges in an image and also in reducing noise.In this filter each pixel in image looks at its adjacent neighbor pixel todecide whether or not it is representative of its surroundings and replace thepixel value with the median of those values. In median filtering,sorting all the pixel value within the window size from the surroundingneighbor into numerical order and then replace the pixel with median intensityvalue of the pixels within the windows. Adaptive median filter5 is an improvedversion of median filter which works within a rectangular region(window) witheach output pixel contains the median value of window neighborhood around thecorresponding pixel in the input images.
This filter is basically used tosmooth the non-repulsive noise from two-dimensional signals without blurringedges and preserved images which makes, it suitable for enhancing mammogramimages. Frost filter 7 is an exponentiallyweighted averaging filter which is used to eliminate the quantum noisefrom the mammograms in which the coefficient ofvariation is the ratio of the local standard deviation to the local mean of thedistorted image. Frost filter uses weighted sum of values with ‘n’ window sizeto replace central pixel and this weighted factor directly proportional to thedifference between central and other pixels.Wavelet transforms filtering have become increasingly important whichplay an extremely crucial role in image processing since it allow both time andfrequency analysis simultaneously.
Wavelet transformdecomposes the input image into four lower sub-bands with approximationcoefficients (LL1) and detail coefficients (HH1, HL1, and LH1). To obtain thenext level of wavelet coefficients, the sub-band LL1 is further decomposed andsampled which results in two-level wavelet decomposition. In this technique,elimination of any of the undesired sub-band or their combinations are doneand then reconstructing the original image using inverse wavelet transform. In our experiments, we use Haar waveletand eliminate HH, LH, and HL bands individually for first- and second-leveldecomposition.Histogram equalization is a nonlinear contrastenhancement which is basically used for enhancing the appearance of images. Histogramequalization enhances the contrast of images by transforming the values in anintensity image, or an indexed image, in such a manner that the output imagehistogram approximately matches a given histogram.(i.
e.it give a lineartrend to the cumulative probability function associated to the image.)For images which contain localregions of low contrast bright or dark regions, histogram equalization doesn’twork effectively so a modified histogram equalization technique called Contrast Limited Adaptive Histogram Equalization can be used on such imagesfor better results which consider only small regions and based on their local cumulative distribution function(cdf), performs contrastenhancement of those regions.
6 The amount of contrast enhancement for someintensity is directly proportional to the slope of the CDF function at thatintensity level. Hence contrast enhancement can be limited by limiting theslope of the CDF. Theslope of CDF at a bin location is determined by the height of the histogram forthat bin.Therefore if we limit the height of the histogram to a certain level we canlimit the slope of the CDF and hence the amount of contrast enhancement. The CLAHE methodseeks to reduce the noise and edges-shadowing effect produced in homogeneousareas and was originally developed for medical imaging.
This method has beenused for enhancement to remove the noise and reduces the edge-shadowing effectin the pre-processing of digital mammogram.Measurement of image quality is very important to manyimage processing systems. Since it inherent physical limitations and economicreasons for the quality of images and videos, such that it could view by ahuman observer. Some of existing measures of image quality are listed below.Imagedifferencing is used todetermine changes between images and Maximum difference(MD) is the maximum of the error between real image and filtered image. Forthe better performance maximum difference should be minimum and large valueshows poor quality.The root mean square error (RMSE) is thesquare root of cumulative squared error between the filtered and the originalimage for comparing various image processing.
For the better performanceroot mean square error shouldbe minimum and large value shows poor quality.The peak signal-to-noise ratio(PSNR) in image processing is astandard measure of the sensitivity of imaging system. For the better performance peak signal-to-noise ratio shouldbe maximum and smaller value shows poor quality.In image processing normalizedabsolute error (NAE) is a measureof difference between the filtered and theoriginal image for comparing various images.
Higher value of normalized absoluteerror is used for poor quality of images.Smaller the value of structuralcontent (SC) better is the image quality and higher value of structuralcontent is used for poor quality of images. For image-processingapplications in which the brightness of the image and template can vary due tolighting and exposure conditions, the images can be first normalized. Normalized Cross Correlation (NCC) showbetter quality with higher value. Laplacianmean squared error (LMSE) is based on the importance of edges measurementwith lower the value of laplacian mean squared error better the quality ofimages. Universal image quality index (IQI) is a image qualitydistortion measurement which is defined as the product of three factors:structural distortion, contrast distortion and luminance distortion, of thedistortion of ideal image with respect to filtered image.
The Structural Similarity (SSIM) indexis a method for measure similarity between ideal image and filtered image. Itis an quality measure of one of the filtered images being compared, provided the ideal image is regarded as ofperfect quality . SSIM is a better than that of universalimage quality index.
The Pratt’s Figure of Merit(PRATT)find the edge location accuracy by the displacement of filtered image from anideal image where both the image (filtered & ideal) are edge detectedimages from the edge detectors like sobel , prewitt , Robert , log etc.The Mammographic Image Analysis Society (MIAS)database is the largest publicly available database of mammographic data. It contains approximately 322 screening mammography cases,where 207 images represent normal, while 64 and 51 images referred as benignand malignant cases respectively.For experimental purpose, the 45 imagesare taken from MIAS database which includes of 15 normal images and 30 abnormalimages. The abnormal images are again classified into two classes which arebenign and malign. There are 15 benign images and 15 malign images.
In this section, the results of those filters are comparedwhich are discussed in previous section. Many proposed work in the literatureare also discussed and compared with the present approach. Furthermore, thefiltered images shown in this section are obtained as a result of subjectingthe breast mammogram shown in Figure1 tovarious filters.The image quality metrics of median filter are presented in Table 1 and the results of medianfilter applied to the original mammogram of Figure1 are shown in Figure 2 . From the image quality metrics table it is found that, theimage quality degrades with the increase in windows size. Some of theperformance of image metrics (MD,NAE,LMSE,RMSE,SC) increase and some (PSNR,IQI,SSMI,NCC) decreases with the increase of windows size .Also pratts figure ofmerit show better performance with lower window size.
Table 2shows the image quality metrics obtained for Adaptive median filter (AMF) and Figure3 shows the results for AMF. In AMF the image quality will not change much withthe increase in windows size .It is found that performance of AMF is quite goodresulting in lower value of RMSE and higher value of PSNR.
The Pratts figure of merit also shows better results as compared tomedian filter.Figure 4 shows the result of frost filter and the image qualitymetrics obtained for frost filter is presented in Table 3. Here the image quality will change variably with theincrease in windows size .It is found that performance of frost filter is quitegood resulting in window size 5.Figure 5 shows the filtered images fromwavelet filtering, and the corresponding tables of image quality metrics arepresented in Table 4. It is found that image qualityis maintained after filtering at first-level decomposition as indicated byRMSE, PSNR, IQI and SSMI while after second-level decomposition, image becomesmuch brighter and Pratt’s figure of merit reduces. When LH band is eliminatedafter second-level decomposition, most of the details are lost giving MD of0.000 and NAE of 254.
00 .The best result is obtained when HH band is eliminatedafter first-level decomposition.The image quality metrics obtained for histogramequalization(column 2) & CLAHE(column 3) is presented in Table 5 and the results for histogram equalization (a)& CLAHE (b)areshown in figure 6. It is found that with maximum difference and psnr inhistogram equalization is smaller than all other filters.In this paper, review and comparison of representative denoisingmethods both qualitatively and quantitatively with extensive experimentsconduct to evaluate the performance of all the algorithms. In analyticalcomparison, it was found that image representations with over complete basisfunctions improve the performance within each category.
In this paper it isclear from the comparison that all the denoising techniques are important forvarious applications. In applications that require high efficiency, somefilters are used, some filters are more appropriate for high searchingcomplexity, memory and complexity issue.