Microcalcification Detection by Morphology, Singularities of Contourlet Transform and Neural Network

-The proposed method presents a new classification approach to microcalcification detection in mammograms using morphology, Contourlet Transform and Artificial Neural Network. Early detection of breast cancer is possible by enhancing microcalcification features obtained using morphology and singularities of Contourlet Transform. The significant edge information indicating the relevant features in various decomposition levels are preserved while removing the artifacts. These features are utilized to detect microcalcifications by classification employing the Back Propagation Neural Network. Target to background contrast ratio, Contrast and Peak Signal to Noise ratio are considered for performance evaluation of the enhancement algorithm. The accuracy of the classification algorithm is 95%. The miniMIAS mammographic database is employed for testing the accuracy of the proposed method and the results are promising. Keywords--Breast Cancer, Back Propagation Neural Network, Contourlet Transform, Morphology


INTRODUCTION
REAST cancer is one of the leading causes of cancer death among women.According to the statistics in 2011-12, the expected mortality rate due to breast cancer is about 39,520 [1], even though there is decrease in death rates since 1990.According to the Indian Council of Medical Research (ICMR), breast cancer becomes the leading cause of cancerrelated mortality among women and it will affect nearly 0.25 million women in India by 2015.Mammography as a screening tool is one of the best proven technique for early breast cancer detection.Mammographic image analysis is a complicated and difficult task which requires opinion of highly trained radiologists.Detection of MiCro-Calcification (MCC), a possible symptom of breast cancer is a complex task because of the inhomogeneous background and the high noise level in the images due to emulsion artifacts.Microcalcifications are tiny, granular, linear, or irregular deposits of Calcium Phosphate Hydroxide which appear as bright white spots with size ranging from 0.1-1.0mm and an average diameter about 0.3 [2].A high contrast is essential in differentiating minute MCC structures with breast tissues.Computer aided image processing techniques assists radiologists in analyzing suspicious areas and to provide a second opinion of diagnosis.The multi scale and multidirectional analysis of the proposed method can effectively enhance sharp variation points indicating the presence of MCC by removing various artifacts thereby increasing the reduction of false positives.

II. LITERATURE SURVEY
Recently, wavelet-based enhancement approaches [3][4] [5] [6] have been employed to acquire better performances.Natural images have higher geometrical characteristics [7], and the discontinuity is always along the smooth curve.The sharp edges and singularities in natural images cannot be efficiently presented by a separable two dimensional wavelet transform even though it provides an optimal representation for one dimensional piece-wise smooth signals.To overcome the drawbacks of Wavelet Transform, Contourlet transform by Minh Do et al [8], an efficient representation is employed to capture the smooth contours that are the dominant feature in natural images.The multi resolution property of Contourlet transform is proficient in separating small objects such as microcalcifications from large objects such as background structures.Manzano et.al [9], proposed an image enhancement technique using an orientation space analysis based on a contourlet transform.Xinsheng Zhang and Hua Xie [10] utilized Contourlet Transform , Generalized Gaussian Mixture Model (GGMM) and Bayesian classifier to enhance the suspicious features.Hu et.al.suggested a detection Algorithm [11] of Suspicious Lesions by Adaptive Thresholding Based on Multi resolution Analysis along with morphological filter to remove the noise and to enhance the gray-level feature and shape feature.Wiselin Jiji.G [12] utilized features from wavelet decomposition and Gabor filter for detecting micro calcification using back propagation neural network.Manimegalai.P [13] developed a system extracting statistical features by wavelet decomposition for classifying breast tissue using Back propagation Neural Network (BPNN).Leena Jasmine [14] used a new approach for micro calcification detection using back propagation neural network and non subsampled Contourlet transform which yielded a significant true detection rate approximately 87%.

III. METHODOLOGY
The proposed algorithm makes use of the multiscale and directionality properties of the Contourlet Transform to  The orientation characteristics of combined impulse response is decided by the directional filters used and the specified number of directional decomposition levels.The major difficulties in identifying MCC in mammograms are due to very high frequency noise and background with low frequency components.To avoid these difficulties, the proposed method utilizes singularities of the multidirectional, multiresolution Contourlet coefficients.The features of an edge are described differently in different directions and at different resolution scales as decided by the level of the Laplacian Pyramid at which the DFB decomposition is performed.Coefficients at various scales are strongly interdependent [18].Po and Do describe the relationship between image features that appear in the contourlet transform at different scales using the parent, child, cousin and neighbour nomenclature.Clinically important microcalcification structures form edge features which are present at more than one level of resolutions in the contourlet decomposition whereas edges due to image artifacts such as those due to Poisson noise are very fine structures that are significant at only the finest resolutions.The parent-child tree is constructed for modulus maxima at each directional subband at each level, in the direction specified by the impulse response of that band.The tree retains only those coefficients that propagate to the coarser levels of the contourlet decompositions, while discarding those that exist only at the finer resolutions.The contourlet coefficients on both sides of and including the selected modulus maxima coefficients are boosted by a scale factor.An inverse contourlet transform which replaces the coarse sub-image by a zero array will yield the locations of possible microcalcifications.Three levels of Laplacian Pyramids, with a four level Directional Filter Bank at each Laplacian Pyramid level, were employed in the proposed method for decomposing the mammogram images in the mini-MIAS database.

Top-Hat Transformation
Enhancement approaches using Morphological operations [19] is of great importance as it is a powerful tool in extracting important information in an image.The two basic operations in morphology are dilation and erosion.Dilation results in growth or object or thickening while erosion shrinks objects based on the structuring element.

{ ( ) }
(1) where I is the image, S is the structuring element and z is the outcome when S is subset or equal to I.

{ [ ( )
] } By combining erosion and dilation, the important morphological filter operations opening and closing are formed.Opening and closing are defined as follows.
( ) Opening operation helps to keep the background (i.e., features that cannot hold the structuring element) which is not required for the proposed method.So a Top-Hat operation [20] is performed to remove the background in order to get the required microcalcification features.Top-Hat representation is given by In Top-Hat operation, the foreground objects can be highlighted by suppressing the dark background.Since the low contrast MCCs in mammograms appear as circularly bright spots, and a calcification has approximately a size of 20 pixels on each mammogram [21], the proposed method considers a structuring element larger than 20 pixels.So the detection of microcalcification is possible through morphological approach.The resulting image is converted to a binary image by thresholding it with the value of 8*σ (empirically obtained), where σ is the standard deviation of the result image.

Proposed Method
The proposed method combines the approach based on the modulus maxima of the contourlet transform and the morphological approach by a logical AND operation to detect clinically important microcalcification structures while removing noise due to emulsion artifacts.The low pass filters employed at the LP stage are derived from the PKVA filters [22] and the fan filters at the DFB stage are derived from the BIOR 9, 7 filters [23].The pixels in the mammogram image corresponding to the location of logic 1 in the image obtained from the previous step are boosted to emphasize the locations of the microcalcifications.To develop a CAD technique that classifies the mammographic images into normal and those with microcalcifications, a Back Propagation Neural Network (BPNN) [24] is employed.For the training set, the ground truth that accompanies the MIAS database [25] is employed.256 x 256 snippets of the mammogram images are employed for training the Back Propagation Neural Network.Five hundred such blocks were employed in the training set.In order to reduce the dimension of the feature vector, the enhanced image is divided into blocks of size 32 x 32.The energy of each block of the denoised and quality enhanced image is computed by squaring every element in the block and adding.In order to simplify the computation, energy value of each block is divided by the maximum energy value of all blocks to get normalized energy values which are less than or equal to one.The lexicographically ordered 64 x 1 energy vector is computed for each mammogram snippet and is considered as the feature vector for the classification process.The BPNN used for the classification thus employs 64 input neurons and one output neuron.Five hidden layer neurons in a single hidden layer were employed for the BPNN.For the test data set 256 x 256 snippets of mammograms similarly enhanced as before, are converted to equivalent 64 x 1 feature vectors.

Algorithm
1. Morphological Operation: Top-Hat operation of the original mammogram image is performed (a 256 x 256 snippet is taken here) to remove the background in order to get the required microcalcification features.

Thresholding:
The resultant image is thresholded using 8*σ, where σ is the standard deviation of the resultant image.

3.
Extraction: Contourlet Transform of original image is computed for three scales with a four level directional decomposition at each scale.The proposed method was tested on mini-MIAS database [25] of digital mammograms by UK research group.The database consists of 322 images of size 1024 x .Am them im .Among these 25 images, diagnosed as malignant and 12 as benign.

V. PERFORMANCE MEASURES
Various performance measures such as Contrast, Peak Signal to Noise Ratio and Target to Background Contrast ratio [26] were considered for measuring th ontrast, C of a region is defined by where f is the mean gray-level value of the foreground and b is the mean gray-level value of the background.PSNR is defined as where p is the maximum gray-level value and σ is the standard deviation of the background.
In order to evaluate the effect of visual appearance of our method, we co e Target to using Variance (TB C ).The expression for computing TB C is nsider th Background Contrast ratio where μ δ is the difference between the ratios of the mean gray levels in the foreground and background and is the ratio o standard deviation of enhanced and original image.

VI. EXPERIMENTAL RESULTS
The mini-MIAS mammographic database was employed or benchmarking the proposed algorithm.Three levels of the lacian Pyramid with four levels of the Directional Filte  with the proposed method indicates that the enhanced image is visually superior with a larger contrast.In order to train the neural network, the images are randomly selected.The proposed method included 3008 blocks of mammographic images for training the network.Validation of the proposed method was conducted by the ground truth and the advice of expert radiologists provided in the database.Proposed method yields 95% true positive rate and 96.71% true negative rate with 96.70% accuracy.Since the false positive rate is very less (3.29%), the proposed method is very helpful reducing unnecessary biopsies.CON orphological operation and selected contourlet coefficient mammographic image enhancement.Topreground microcalcification fe band coefficients obtained from contourlet transform on either side of the selected coefficients in the corresponding tree were boosted while those that exist at only the finest scales were suppressed.Relevant edge features including microcalcification structures were enhanced while image [20] F. Meyer, "Contrast feature extraction artifacts that exist at the finest scales were suppressed.This method assists radiologists in providing a second opinion for detecting breast cancer in its early stage.Quality measures including contrast, PSNR and TBc were calculated to determine the efficiency of the proposed method.Experimental results showed that the resultant enhanced images using the proposed method are effective in improving the contrast and visual appearance by reducing artifacts.Accuracy of the proposed method is found to be 96.7%.
subbands at each level of resolution in the specified directions are computed.5. Artifact reduction: To re rent-child relationships from finer to coarser levels are retained and others are pruned out.6. Boosting: The modulus maxima that are retained after step 5 and the coefficients 7. Coarse suppression: Assign zero arrays to the coarse subband.Apply Pyramidal on on the selected coefficients in the directional subbands.8. Thresholding: Threshold the resultant image using the value 8*σ, where σ is the standard deviat 9. D ing: Apply a logical AND operation with the results obtained from step 2 and step 8 to locate the MCCs.10. cement: The pixels in the mammogram image corresponding to the location of logic 1 in the image obtained from the previous step locations of the microcalcifications.11.Feature Reduction: Divide the enhanced image into blocks of size 32 x 32 to reduce the size of the feature vector.12. Feature Vector generation alize it to generate the 64 x 1 feature vector.13.Classification: Train the BPNN Each element (256 x 256 mammogram snippet) is similarly converted to a 64 x 1 energy vector as before, prior to presentation to the BPNN for classificatio IV.DA
im s.The PKVA filters were employed for the Laplacian Pyramid and BIOR 9, 7 filters were employed at the DFB stage.Figure4shows the results of the algorithm on the image MDB 241 from the mini-MIAS database, for the major intermediate steps of the proposed algorithm.The enhanced results for the Mdb241, Mdb211, Mdb245 and Mdb249 images from the mini MIAS database are shown in figure5.Mdb249 is a mammographic image having dense glandular background tissue with welldefined malignant microcalcifications.Mdb211 is a fatty glandular mammogram with a difficult to detect microcalcification cluster.Mdb245 represents a mammographic image with widely distributed calcifications.Mdb241 is a mammographic image having easy to detect microcalcifications.The results in figure5reveal that the proposed technique enhances the appearance of the microcalcifications against the surrounding dense tissue which may otherwise obscure these structures.
2001.Her area of research interest is Image Processing.She has 4 papers in international journals 2 papers in national conferences and 4 papers in international credit.Currently she is working as Assistant Professor at g College, Aluva, Ernakulam.VinuThomas is at present, working as Associate Professor in Govt.College of Engineering, Cherthala, managed by the Institute of Human Resource Development, Govt. of Kerala, India.He was awarded PhD in Electronics (Develop conferences to her K.M.E.A Engineer University of Science and Technology (CUSAT) in 2009.He secured the Gold Medal and first rank for M-Tech in Electronics from the Department of Electronics, CUSAT, for 1999-2001.He did his B-Tech degree in munication Engineering from Mar Athanasius College of angalam, Kerala, during 1989-1993.He has 18 papers in s, 11 papers in international conferences and 5 papers in to his credit.Electronics and Com Engineering, Kotham international journa national conference