Breast Ultrasound Images Enhancement Using Gray Level Co-Occurrence Matrices Quantizing Technique

This article demonstrates a simple and robust enhancement method for breast ultrasound images based on quantizing the gray level intensities. The quantizing is performed using gray level co-occurrence matrices; that calculates the neighbor intensity interrelation according to the number of gray intensities per level. In this research we divide the gray scale to 15 levels. Gaussian and median filtrations were implemented and iterated 7 times, at each level, using a kernel size of 11x11. Finally each filtered level is translated back to its original location. This quantization technique significantly smoothes the breast ultrasound image while preserving edges. The performance of the algorithm has been compared with the standard filtering technique and evaluated using second order statistical methods. Test and synthesized images with induced speckle noise were used for technique verification and automatic edge detection. The proposed method demonstrates high filtration quality performance and edge preservation compared to the standard overall image filtration method. The textures were preserved with slight blurring. The proposed method introduces a new enhancing technique based on second order dependency matrices quantizing technique.


Introduction
Breast cancer is one of lead ing death in wo men [1]. Ultras ound imaging modality has a major ro le in the diagnosing of breast abnormalit ies due to its flexib ility, safety and low cost compared with mammography. Breast ultrasound scanning is reco mmended rather than mammography for early detection of young females [1]. Ultrasound scanning is beset with noise such as speckle noise due the nature of ultrasound and the physics of interaction with tissues. The objective of speckle reduction is image enhancement and proper segmentation. Image noise cancellation algorithms encountered with some drawbacks such as reduction of texture details and edge blurring. Gray level co-occurrence mat rices (GLCM) calculate the pixel interrelation based on particular distance in the four major directions. Several texture features extracted fro m the GLCM such as contrast, ho mogeneity, co rrelat ion and energy. Most of the time, researchers use GLCM texture feature descriptors as analysis and classificat ion tools. A Mohd et al use GLCM to classify masses in mammo grams [2]. Hari W. et al imp lement GLCM and Gray-level and Run Length Matrix to classify cyst and Non-cyst in

Method and Implementation
The proposed algorith m has been developed, in MATLAB environment, and tested using registered breast ultrasound fro m CD database [8]. Further verification of the method was done using synthesis images corrupted with noise. The sequence of image processing consists of three major stages: i-quantizing stage in which the original image is quantized into 15 correlated intensity levels ii-filtrat ion and 7-times iteration stage that uses Median and Gaussian filters, and iiireconstructing the image by translating each filtered quantized sub image to its orig inal location.

Gray Level Quantizing, Using (GLCM)
The Gray level co-occurrence matrix (GLCM ) co mputes the probability density function of image f(x,y) for all pair of pixels (i) and (j) in d istance (d) with angular displacement (θ) =0,45, 90 and 135 degree [9], the calculat ion will co mpute the frequency of gray tone occurrence for angular adjacent pixels. The GLCM has been used for two purposes, quantizing the ultrasound images to improve the filtrat ion, and extraction of features' descriptors. In term of theory quantising or gray level slicing amounts to a mult iplication by a window of a specific size in space domain. Such mu ltip licat ion will translate high frequency components (edges) to baseband frequencies (low frequencies). The Median and Gaussian filters (second stage) work very well at these low frequencies.

Filtrati on and Iteration Stage
Several techniques for despeckling of u ltrasound images have been documented [10]. The minor reflect ions of ultrasound wave develop the speckle noise wh ich reduces the contrast and the small details delectability [10]. A linear Gaussian filter and a non-linear median filter were implemented to smooth the images, the size of the filters was 11x11 [9]. Linear isotropic Gaussian filter is common ly used to smooth the image. A Median filter calcu lates the med ian value of the moving window and replaces it with the central window value. Median filters are effective in removing step noise [11]. The utilizing and limitation of such filters in image processing has been demonstrated[10] [11]. A Linear and non linear techniques were used for noise reduction, enhancement the contrast and preserve the edges. The proposed standard filt ration techniques were iterated seven times to reduce the image noise. The iteration has no blurring effect, though it reduces noise. The filtering of each quantized level has resulted in good image filtration process with min imal edge blurring.

Translati on and Image Reconstruction Stage
The original image is sliced into 15 quantized frames; each frame has the same size of the original image containing interrelated of 17 gray tones co-occurrences. Each frame is filtered 7 t imes and translated back to the original image location to reconstruct the filtered image.

Results
The nature of ultrasound interaction with tissue develop artefact, low contrast and low resolution image. The overall mechanis m of the proposed filtration technique performs the low pass filtrat ion in special manner, where the low frequency components are well filtered while preserving the high frequency components; this technique keeps the edge highlighted. In literature it is preferable to imp lement small size kernel lo w pass filter in order to preserve the fine details and image textures from bullring, our proposed technique is able to use large kernel size with less over smoothing results.
Ultrasound breast image database CD is used to evaluate our proposed method. The gray quantization introduces a special technique that preserves edges; the edges are spared in different filtered levels. The proposed method is ab le to reduce the noise and preserve the image local and g lobal details, wh ich is important step in breast lesion diagnosing and classification. Figure (1) shows an example of a noisy breast ultrasound image processed by the general and the quantized filtration technique. (a) is original breast image (ORG.), (b) corrupted with speckle noise of 0.01 variance, (c) the proposed quantizing method (QZ.) result, and (d) is the standard general filtrat ion (GRN) output. Visual inspection indicates that, the proposed filtration technique smooths the homogenous regions effectively and preserves the major edges. Another examp le of breast image processing is given in Figures (2) and (3). The 15 levels of quantization are found to be suitable given the limited resolution of ultrasound imaging.

Evaluation
The processed image has been evaluated using visual perception and second order statistics that estimate the overall imp rovement in the texture. The GLCM features descriptors used in our analysis are: contrast (Cont), correlation (Corr), ho mogeneity (Ho mo ) and energy (Enrg). The descriptors' equations are given in Tab le (1), where C is the concurrence of gray tone at pixel (i) and pixel (j) [7] [9]. The contrast descriptor which is presents the power of the intensity similarity interrelation of the image. The homogeneity feature shows progressed flatness degree of the division of co occurrence elements. The correlation descriptor shows improve lin ked interpixel redundancy value. The energy represents enhancement in the power of co occurrence.
The evaluation was done to statistically estimate the performance (P) of the proposed method in comparison to the standard general filtration technique (GRN.). Table (2) presents the second order GLCM statistics, the proposed algorith m has remarkable texture similarity to the original image features, the interrelation is (-2.3%). The ho mogeneity feature shows progressed flatness co occurrence elements degree (2.99%). The correlat ion descriptor shows improve lin ked interpixel redundancy value (10.95%) and the energy represents enhancement in the power of co occurrence (2.47%). The interpretation of the GLCM features is that the low descriptors values mean high texture similarity to the original image at least 98.05% similar. Co mpared to the general filtration technique, the intensity interrelat ion is highly degraded (-88%), see Table (2). While the over smoothed image increases the homogeneity (45.8%), boosting the image correlation (142%) and energy descriptor (35%).  A synthetic image is used to estimate the performance of our proposed algorithm. Evaluation tools such as visual perception, N. Otus's automatic edge detection, statistical analysis and image texture, were imp lemented to assess the effect of the algorithm in local and global details. Figure (4) shows a phantom image (a), corrupted with speckle noise of variance of 0.01(b), Then filtered with our method (c) and the general method technique in (d). Reg ion of interest (ROI) has been cropped as in figure (5), enhancement in contrast is observed. Figure (6) presents a cross section of the synthesized image (a) orig inal image (b) the quantized filtration (c) the general filtration, the result shows good image restoration with the proposed method but the general method blurred some details such as edges. Figure (7) shows the implementation of auto matic N. Otsu's edge detection of the original image (a), the proposed method effectively highlighted the edge (b), mo re than the general technique in (c).
Statistical evaluation has been imp lemented to estimate the effect of the proposed method on GLCM textu re descriptors presented in table (3), the proposed method adequately preserve the texture features of the syntheses image, the worst descriptors is around 16%. In other hand the common filtrat ion dramat ically deteriorate the textures features.    Gray slicing technique (GS) wh ich is based on direct threshold selection of scaled gray tone levels has been implemented as a control in the evaluation of the quantizing technique. Figure (8) shows a region of interest cropped fro m the ultrasound synthesized image (top) and the corresponding filt rated images fro m left the quantizing technique and gray slicing technique respectively, figure (9) the corresponding automatic edge detection .The visual perception shows better noise reduction and texture features preserving when GLCM quantization is used. Table (4) presents the statistical evaluation of the gray slicing and quantizing filtration, the overall performance (P) of the gray quantizing looks better than the gray slicing technique.

Conclusions
Gray tone quantizing technique has been proposed to enhance the filtration performance and preserve the image texture. Quantizing the image into 15 levels proved to be adequate for ultrasound scan. Registered breast ultrasound image data base and synthetic images were used to implement the proposed method. Gray threshold technique or gray slicing results were co mpared to the quantization technique performance. Second order statistic calcu lations have been implemented to evaluate the image textu re features. The proposed method of the gray quantizing technique produces novel results. The proposed method establishes a new technique for edge preservation. The proposed method effectively reduces the noise, preserves the texture and edges. Future work will include the analysis of the method in the space and frequency domain and the effect of the slice thickness on the image enhancement.