Mammogram Restoration under Impulsive Noises using Peer Group-Fuzzy Non-Linear Diffusion Filter

Transcription

1 International Journal for Science and Emerging ISSN No. (Online): Technologies with Latest Trends 22(1): 41-46(2017) ISSN No. (Print): Mammogram Restoration under Impulsive Noises using Peer Group-Fuzzy Non-Linear Diffusion Filter Rajesh Kochher*, Dr.Gagan Marken** and Dr.Anshu Oberoi*** *Research Scholar, Department of Computer Science NIILM University Haryana, India **Assistant Professor, Department of Computer Science NIILM University Haryana, India ***HOD, Department of Computer Science, I.K.G Punjab Technical University Campus Hoshiarpur, India Abstract Breast Cancer is the most widespread cause of deaths among women and determination of abnormalities is a very difficult task. The widely technique used for diagnosis of breast cancer is Mammography, where classification of masses is done using gray weighted function related to HSV color space based on statistical features. Due to the growing demand of digital mammography images and its processing services over communication networks with precise measurement of the quality of mammography images to offer the best quality increases research attention in recent years. The major issue of quality degradation of the mammography is noise introduced during the capturing or transmission of the image. Therefore, developing accurate noise removal algorithms is a key requirement to improve its QoS in digital image processing applications. In this paper, we analyzed different noises models on mammographic images and noise removal techniques independently, and investigate relevant issues based on different quality metrics. Experiments have been conducted for subjective analysis of results and comparing the performance of the quality metrics. Index Terms Breast Cancer, Mammogram masses, Weight function, Salt &- Pepper Noise, Gaussian Noise, Wiener Filter, Order Filter I. INTRODUCTION The Breast Cancer is considered to be one of the widespread causes of death among women and second highest cause of deaths among humans. It has been predicted that each single minute a woman is interpreted with breast cancer and every 13 minutes a woman expires due to disease. Many practices and systems had been introduced for early recognition and diagnosis of breast cancer, but the mammography is the most acquiescent method for early prophecy of disease chiefly through espial of masses and micro calcification [1,2]. Medical experts read mammograms and try to identify the abnormalities present in the breast mass. It has been predicted that 10-30% of women interpret with breast cancer have false negative mammograms and contributed radiologist failure to identify the breast cancer at an early stage due to misinterpretation or noise. According to the medical point of view, reading and interpretation mammograms are the very complex thing [3]. In recent years, computer -aided diagnosis methods are used to guide physicians in early diagnosis and detection of breast cancer, where Computer aided diagnostic tools are used to determine the abnormalities present in the mammogram such as masses, calcification and with the help of these results experts can diagnose and detect breast cancer at an early stage [4]. Computeraided techniques are used for pre-processing and evaluation of images as supporting or second reading The first stage of the computer aided diagnostic tool is to capability to determine abnormalities in breast and the second stage is to diagnose the abnormalities found in masses identified in first stage. The most important step before implementing two stages of computer- aided diagnostic tool is pre-processing stage has to take place that is segmentation of breast part from the whole background and removal of noise components from it [2]. Most of the recent works done on mammography are almost based on mammography s histogram and it has been found that histogram based on the mammograms are not much effective for classifying the breast masses. This is due to the fact that histograms based on mammograms pattern changes mainly due to noise and over - intensification of mammogram images [1].

2 42. Rajesh Kochher* Dr.Gagan Marken** and Dr.Anshu Oberoi*** Noise is the random variation of spectral information and generally regarded as an undesirable by-product of image capture and can be categories in different classes based on their cause of occurrence such Impulse noise, Gaussian noise, Shot noise, Quantization noise, Film grain, Multiplicative noise (Speckle noise) and Periodic noise [5]. The prominence in this research paper is concentrated on experimental evaluation of Salt-and-pepper Noise, Gaussian Noise, Quantization Noise, and Sparkle Noise in Mammographic pre-processing as digital image are mostly affected with those noises [6]. Also, algorithms used to remove this noise are analyzed and compared based on subjective observations. Mathematically, Noise is represented by the percentage of the degraded pixels from the original image and expressed as [7]: In above Equation, digital image M (i,j) is represents as a sum of N(i,j) &- O(i,j), where N (i,j) represents the noisy pixel & O(i,j) represents the noise free pixel. Many active research projects concerned with Noise removal or reduction from digital image processing are going on in the field of Digital Image Processing. In this paper, we have analyzed the effect of different noise models during preprocessing stage on mammographic images and examined different noise removal techniques independently, and investigate their relevant issues based on different quality metrics. The rest of the paper is organized as fallow, Section II describes the different sources of noise in digital image processing and their techniques as well as different noise removal algorithms, and Section III describes brief literature survey on research done on different noise sources and their removal algorithms till date. In section IV these noise models are implemented on a digital image and Noise removal algorithms are applied for subjective analysis and Section V conclude the paper. II. NOISE IN IMAGE &- REMOVAL ALGORITHM Mammographic Image is corrupted by the noise during the transmission and the acquisition which degrades image due to miss-focus of lens, atmospheric turbulence, relative motion between camera and object causes the motion blur. The different kind of noises that appears in mammographic images modeled mathematically are discussed below. A. Gaussian Noise: Gaussian Noise has a probability density function of the normal distribution and Probability density function is given as: Where z represents grey level, µ is average value of z, σ is standard deviation, σ 2 Is variance [5]. Figure 1: PDF of Gaussian Noise Figure 2: PDF of Salt and Pepper Noise B. Salt &- Pepper Noise: Salt and Pepper Noise In the grey scale images salt & pepper noise is caused by memory fault locations or there can be timing errors in the process of digitization [8]. In this noise there are two assumptions of two saturation values such as a and band they are equal to the maximum and minimum values of the digitized image and the probability of each value is less then 0.2.If the value of each probability exceeds the noise will swamp out image [9].This noise is also called as impulse noise is given as Impulse. Where C. Quantization Noise: Quantization Noise is also called as uniform noise. This noise is caused by

3 43. Rajesh Kochher* Dr.Gagan Marken** and Dr.Anshu Oberoi*** quantizing the pixel of image to a number of discrete III. SYSTEM MODELLING levels because of uniform distribution [10].It is given In this section, various noise models are implemented by: on mammographic image database, obtained from Digital Database for Screening Mammogram (DDSM) to conduct experiment. The mammogram s content can be accomplished by using four levels of gray. While describing the masses in mammogram they are D. Speckle Noise: Speckle Noise is also called as enclosed with pixels with slight fluctuations in gray granular noise [10] and multiplicative noise that levels and configure smooth boundaries. Thus, it is mainly occurs in all coherent imaging system like a very necessary to catch the boundary representation laser, acoustic, SAR (Synthetic Aperture Radar) of the pixel values for segmentation of breast masses. imagery. This noise follows a gamma distribution The approach is conducted on spatial domain image and is given as processing with the use of HSV colour space properties and this color space is early relevant to human perception colors and for every pixel a weighted value is determined using a gray weighted function which seizes the gray part of the pixel and is The goal of de-noising is the removal of the noise very vigorous against noise and other noise while retaining the important signal features of parameters. It is given as: original image as much as possible. De-noising is done through the filtering method can be linear filtering or non linear filtering. The prominence in this research paper is concentrated on the study, evaluation and performance comparison between BFED and PGFND. A. BFED is a non-linear filter. It is more suitable for noise reducing for mammographic images. It works on the basis of intensity value of every pixel of images which are connected to other pixel and form a complete image. This filter includes the weight of image in different environment such as color intensity and depth distance of pixels. B. Peer Group-Fuzzy Non-linear Diffusion Filter (PGFND) is the combination of PGFM and NDF method. In PGFND algorithm, the sequence of functioning is implementation of PGFM followed by NDF. The stare set with fuzzy metric algorithm removes the impulsive noise and the gaussian noise is eliminated by NDF and both methods to eliminate speckle noise. Where Peer Group Fuzzy Metric (PGFMA) denoising algorithm works in two steps, first rung detects erroneous pixels and the subsequent is used to precise them [11], [12] given by: Non-linear diffusive filter (NDF) As mentioned in the introduction, a class of image restoration methods is based on the use of non-linear diffusion appear associated to variation problem and may be obtained from the minimization of the appropriate functional and choice of a particular functional depends upon the specific goal of interest [13], [14]. Figure 5: Partial Derivative Variation of Wgray(S, I) with different Intensities The extensity of (S, I) is (0-1) and it evaluates the gray portion of the pixel by using together the intensity and saturation values. In-order to seize the gray content of the pixel, closely agitate the either of the values of R, G or B, which effects the value of saturation of pixel. The gray weighted function is adjustable with saturation values and is found to be stable. It is very clear from equation1 that r1 must take value nearly higher than zero and r2 must hold value lesser than 1 for having stable and continuous gray weighted values. From recent works, it had been predicted that r1=0.1 and r2= 0.85.

4 44. Rajesh Kochher* Dr.Gagan Marken** and Dr.Anshu Oberoi*** Table 1.Statistical Features of Extracted Mass Features Area Mathematical Representation n = Number of pixels Standard Deviation Mean Entropy Variance The mathematical representation of the various features such as standard deviation, varience, entropy and mean extracted from the mammogram are represented in tabular form as shown in Table 1. VI. EXPERIMENTAL IMPLENTATION & DISCUSSIONS To measure the quality of restoration of denoised resultant mammographic image, we have two estimation techniques that are subjective and objective evaluation, where subjective assessment is awkward, time-consuming and exclusive. Lately, lots of efforts have been done to develop objective image quality metrics [15]. In this research full concentration is given on objective quality metric like MSE, PSNR, and SSIM. Mean Square Error (MSE) in digital image processing is calculated by averaging the squared intensity of the original (input) image and the resultant (output) image pixels, mathematically representation as: Peak Signal-to-Noise Ratio (PSNR) or Signal-tonoise ratio (SNR) is a mathematical gauge of image quality depends on the pixel difference between two images [16]. The SNR measure is an approximation of quality of reconstructed image compared with original image. PSNR is defined as Where s = 255 for an 8-bit image. The PSNR is basically the SNR when all pixel values are equal to the maximum possible value. The Structural SIMilarity (SSIM) index is a method for measuring the similarity between two images. The SSIM index can be viewed as a quality measure of one of the images being compared, provided the other image is regarded as of perfect quality. Figure 2: Gaussian Noise Mammogram with varying S.D (σ) values Original mammographic images are converted in grayscale with intensity of pixels varies from 0 to 255. The test mammograms are engendered by accumulating gaussian, impulsive and speckle noise to the image with the functions of MATLAB. The mammogram image by means of Gaussian noise with varying standard deviation (σ) values is shown in figure 2, which depicts that with increase in σ the mammogram becomes more blur and noisy. Figure 3: BFED based Denoised Gaussian Noise Mammogram with varying S.D (σ) Figure 4: PGFND based Denoised Gaussian Noise Mammogram with varying S.D (σ) Applying the BFED and PGFND filters to the image with 10% fixed Gaussian Noise as shown in figure 2, it is observed subjectively the quality of the filtered mammogram with PGFND filtering is better as compared to BFED as shown in figure 2 &3. The objective observations are summarized in table 2. From the tabular results it is observed that MSE performance of BFED is slightly better than PGFND under Gaussian Noise, but with increase in standard deviation BFED performance degrades and PGFND performs better as compare to BFED filtering technique. Further under PGFND technique MSE value remains almost constant for all values of σ, where in BFED MSE increases with increase in standard deviation. Table 2: Mammogram Denoising Results for Gaussian Noise under varying σ Filtering Technique MSE PSNR SSIM σ = 10 BFED PGFND σ = σ = As discussed above, another mammogram is taken from the dataset and salt & pepper noise with Speckle noise 10% fixed) is introduced to it using MATLAB

5 45. Rajesh Kochher* Dr.Gagan Marken** and Dr.Anshu Oberoi*** functions as shown in figure 5(a). To noisy noising the mammogram the blurring effect is far less mammogram BFED and PGFND filters are applied, in PGFND implemented mammogram. Moreover the the resultant mammograms are shown in figure 5(b) picture quality of the filtered mammogram with & 5(c). From subjective observations of figure 5(b,c), PGFND filtering is for better as compared to BFED. the quality of the filtered mammogram with PGFND From the objective observations tabular results in filtering is for better as compared to BFED. The table 4, mean square error decreases with increase of objective observations are summarized in table 3. standard deviation for both the filtering techniques, From the tabular results, it is observed that BFED though the MSE effect is higher in case of PGFND as performed better than PGFND for Mean Square compare to BFED filtering technique. Further form Error, when salt & pepper noise is introduced with tabular results of table 5, PGFND filtering technique speckles. But on the other hand PGFND filtering perform better than BFED for all values of σ. With perform for better than BFED in terms of PSNR and increase in standard deviation (σ), BFED filtering SSIM for all values of σ, Further BFED filtering performance almost remains almost constant for performance degrades with increase in standard PNSR & SSIM, where PSNR & SSIM decreases with deviation as shown in Table 3. increase in σ for PGFND. Figure 5(a): Salt & Pepper with Speckle Noise Mammogram with varying S.D (σ) values Figure 6(a): Gaussian, Salt & Pepper, Speckle Noise Mammogram with varying S.D (σ) values Figur 5(b): BFED based Denoised Salt & Pepper with Speckle Noise Mammogram with varying (σ) Figure 6(b): BFED Denoised Gaussian, Salt & Pepper, Speckle Noise Mammogram with varying S.D (σ) Figure 5(c): PGFND based Denoised Salt & Pepper with Speckle Noise Mammogram with varying (σ) Table 3: Mammogram Denoising Results for Salt & Pepper with Speckle Noise under varying σ Figure 6(c): PGFND Denoised Gaussian, Salt & Pepper, Speckle Noise Mammogram with varying S.D (σ) Table 3: Mammogram Denoising Results for Salt & Pepper with Speckle Noise under varying σ Further all the three noise functions (Salt &Pepper, Gaussian and Speckle Noise) are introduced in a mammogram as shown in figure 6(a) using MATLAB functions as shown in figure. To noisy mammogram BFED and PGFND filters are applied as noise removal algorithms, the resultant mammograms are shown in figure 6(b) & 6(c). From subjective observations of figure 6(b,c), the after V. CONLUSION This research is based noise removal PGFND and BFED algorithms to demonstrate the comparative impact of these filtering techniques on mammogram images using Gaussian, salt & pepper noise and

6 46. Rajesh Kochher* Dr.Gagan Marken** and Dr.Anshu Oberoi*** speckle noise models. The comparative analysis for [15] Wang C. A., "Image quality assessment: from error visibility to structural similarity," IEEE Trans. Image random behavior of these filtering techniques is Processing, vol. 13, no. 4, pp , observed using application-oriented metrics such as [16] Jean-Bernard M., "Image dissimilarity," Signal MSE, SSIM and PSNR. PGFND outperforms BFED Processing, vol. 70, no. 3, pp , filtering technique in terms of PSNR and SSIM under noise models discussed above. With small standard deviation (σ) both filtering techniques perform almost same, but BFED performance starts degrading with increase in σ. In our experimental results, PGFND outperforms BFED and is the best suited noise removal technique for mammogram images. Where MSE both techniques have almost same results under used parameters. [17] J.C. Russ, The image processing Handbook, CRC Press, REFERENCES [1] Surendiran B., Vadivel A., Classifying Malignant and Benign Masses using Statistical Measures, International Journal on Image and Video Processing, vol. 02, pp , [2] Abbadi N., Taee E., Breast Cancer Diagnosis by CAD, International Journal of Computer Applications, vol.05, pp , [3] Khezri R., Hosseini R., Mazinani M., A Fuzzy Rulebased Expert system for the Prognosis of Risk Development of Breast Cancer, International Journal of Engineering (IJE), vol.27, pp , [4] Tang J., Ranyayyan R., Computer Aided Detection and Diagnosis of Breast Cancer with Mammography: Recent Advances, Proceedings of IEEE on Information Technology on Biomedical, vol. 2, pp , [5] Garg R., Kumar A., Comparison of Various Noise Removals using Bayesian Framework, International journal of modern engineering, Vol.2, Issue.1, [6] Ishta, Goel A., Gupta R., Fuzzy based new algorithm for noise removal and Edge Detection, IJARCST, vol. 2, pp , [7] Kumar P., Kumar R., Comparison of adaptive mean filter based on homogeneity level information and the new generation filters, IOSR Journal of Computer Engineering, Vol.14, pp , [8] Bansal.R., Sehgal P., Bedi P., A Simplified Fuzzy Filter for Impulse Noise Removal using Thresholding, Proceedings of IEEE, San Francisco, USA, [9] Kundra H. Verma M., Aashima, Filter for Removal of Impulse Noise by Using Fuzzy Logic, International Journal of Image Processing (IJIP), Vol. 3, issue 5, [10] Farooque M., Rohankar J. S., Survey on Various Noises and Techniques for De-noising the Color Image, International Journal of Application/ Innovation in Engineering & Management (IJAIEM)), vol. 2, issue 11, pp , [11] Camarena J., Gregori V., Morillas S., Sapena A., Fast detection and removal of impulsive noise using peer group and fuzzy metrics, Journal of Visual Communication and Image Representation, vol. 19, pp. 20, [12] Smolka B., Fast detection and impulsive noise removal in color Images. Real-Time Imaging vol. 11, pp.389, [13] Rudin L.I., Osher S., Fatemi E., Nonlinear total variation based noise removal algorithm, Physica D, vol. 60, pp. 259, [14] Vogel C., Oman E., Iterative methods for total variation denoising, SIAM J. Sci. Comput., vol. 17, 1, pp. 227, 1996.