Recognition Of Abnormalities In Digital Mammograms Biology Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

David J. Marchette et. al., worked on the fundamental problem of automating the detection and recognition of abnormalities in digital mammograms utilizing computational statistics is one of extracting the appropriate features for use in a classification system. They considered a method of using boundaries to segment the window into more homogeneous regions for use in the feature extraction calculation. This approach has been applied to the problem of discriminating between tumor and healthy tissue in digital mammography. In order to evaluate the utility of the features for classification they constructed the probability density functions for the features associated with each class. These can be used for classification by the standard likelihood ratio test and so can give a good measure of the utility of each individual feature. Experience has shown that these densities are not well modeled by any of the more common families of densities, and so a nonparametric approach is taken. This is also the reason for using density estimation techniques rather than linear or quadratic classifiers.

Dimitris K. Iakovidis et. al., [2] in 2006 proposed a novel scheme for efficient content-based medical image retrieval, formalized according to the PANDA (PAtterns for Next generation Database systems) framework. The proposed scheme involves low-level feature extraction from image regions followed by clustering of the feature space to form higher-level patterns. The components of each pattern include a cluster representation and a measure that quantifies the quality of the image content representation achieved by the pattern. The similarity between two patterns is estimated as a function of the similarity between both the structure and the measure components of the patterns. Experiments were performed on a reference set of radiographic images, using standard wavelet domain image features.

H.S. Sheshadri, [3] in 2006 proposed a novel approach for the development of a Computer Aided Diagnosis (CAD) system for mammogram image analysis. The authors have made an attempt to implement a CAD system for the detection of microcalcification from mammogram image segmentation and analysis. They presented a new method which improves the classification performance of the co-occurrence approach. In order to achieve they combined the co-occurrence approach and one of the filtering approach. The images are segmented using simple thresholding technique.

Nuryanti Mohd. Salleh et. al. [4], in 2008 evaluated proposed morphological features to classify breast cancer cells. The morphological features were evaluated using neural networks. The features were presented to several neural networks architecture to investigate the most suitable neural network type for classifying the features effectively. The performance of the networks was compared based on resulted mean squared error, accuracy, false positive, false negative, sensitivity and specificity. The optimum network for classification of breast cancer cells was found using Hybrid Multilayer Perceptron (HMLP) network. The HMLP network was then employed to investigate the diagnostic capability of the features individually and in combination. The features were found to have important diagnostic capabilities. Training the network with a larger number of dominant morphological features was found to significantly increase the diagnostic capabilities. A combination of the proposed features gave the highest accuracy of 96%.

Matteo Roffilli [5], in 2006 aimed to make some contribution to the demonstration of the applicability of Machine Learning technologies to the diagnostic problem. The problem chosen for this work is the diagnosis of breast tumour through digital mammography. The reason for this choice lies in the fact that this problem is regarded as one of the hardest to solve on the field of objects recognition. This thesis specifically proposes a cancer detection system which is capable to find lesions not by means of an algorithm based on prearranged features, but by analogy with other previously analyzed lesions. In other words, the system is trained to recognize different typologies of lesions which are present on a dataset prepared for the purpose. The system can also find autonomously lesions with similar characteristics in unknown mammograms. This innovative achievement is made possible by the use of SVM as classifier for classification.

Akram I. Omara et. al., [6] in 2006 developed a Computer Diagnosis System as an aid to radiologists, which could be helpful in diagnosing abnormalities faster than traditional screening program without the drawback attribute to human factors. The techniques used for feature extraction is based on the wavelet decomposition of locally processed image (region of interest). Both the wavelet coefficients and the statistical measures of different wavelet detail levels are used as features that describe effectively any normal and abnormal region. They used two techniques for the classification stage: the minimum distance classifier and the voting K-Nearest Neighbor classifier.

Yih-Chih Chiou, [7] in 2006 presented an innovative algorithm for the automatic registration of mammograms. The proposed algorithm uses a feature-based registration method to match internal features and an intensity-based registration method to match external features. In this research, we used three approaches to derive transformation parameters, i.e., internal matched pairs only, external matched pairs only, and both internal and external matched pair. The mammograms were obtained from the MIAS digital mammogram database. The results suggest that both internal and external features should be used in registration if possible.

Michal Haindl, [8] in 2008 presented a multiscale unsupervised segmenter for automatic detection of potentially cancerous regions of interest containing fibroglandular tissue in digital screening mammography. The mammogram tissue textures are locally represented by four causal multispectral random field models recursively evaluated for each pixel and several scales. The segmentation part of the algorithm is based on the underlying Gaussian mixture model and starts with an over segmented initial estimation which is adaptively modified until the optimal number of homogeneous mammogram segments is reached. The performance of the presented method has been verified on the Digital Database for Screening Mammography (DDSM) from the University of South Florida as well as extensively tested on the Prague Texture Segmentation Benchmark and compares favorably with several alternative unsupervised texture segmentation methods.

K.Thangavel, [9] in 2005 proposed a method for detection of microcalcification based on textural image segmentation. Markov Random Field method hybrid with Ant Colony System, Genetic Algorithm and Backpropagation Network (MRF-ACSGA-BPN) is implemented for the detection of microcalcification in digital mammogram. Identification of microcalcification is performed in two steps namely, segmentation and classification. First, the mammogram image is segmented using MRF-ACSGA method to extract the suspicious region. Second, the conventional textural analysis methods such as Spatial Gray Level Dependency Method (SGLDM), Surrounding Region Dependence Method (SRDM), Gray-Level Run-Length Method (GLRLM) and Gray Level Difference Method (GLDM) are used to extract the features from the segmented image. A three-layer Backpropagation Neural Network classifier, trained by jack knife method and round robin method, is used to classify the extracted features into benign or malignant. The classification performances for the texture-analysis methods are evaluated using a Receiver Operating-Characteristics (ROC) analysis. The proposed algorithm and the techniques are tested on 161 pairs of digitized mammograms from Mammographic Image Analysis Society (MIAS) database.

Arianna Mencattini,, [10] in 2008 illustrated a method to classify clusters of microcalcifications characterizing the lesion by the extraction of geometrical (2D) and textural (3D) features. Then, through a statistical analysis of these features, we can choose the most discriminating between benign and malignant lesions and so design the classifier. Specific aspects of contrast enhancement and features selection have been described, and preliminary results, obtained on mammographic images, are provided.

H. S. Sheshadri, [11] in 2005 presented a new statistical algorithm which partitions a mammogram into homogeneous texture regions. The algorithm assigns each pixel in the mammogram membership to one of an eight number of classes depending upon the statistical properties of the pixel and its neighbors. The individual pixel classifications form a two-dimensional labeled which must be estimated from the observed image. Both the mammogram and its labeled fields are modeled as discrete-parameter random fields. They estimated the pixel classes by minimizing the expected value of the number of misclassified pixels; this is known as the Maximizer of the Posterior Marginals (MPM) estimate. The Expectation-Maximization (EM) algorithm is employed to estimate from the observed mammogram the unknown parameters needed for the MPM estimate.

T.Balakumaran,[12] in 2011 presented a novel approach for Microcalcification Cluster Detection in Digital Mammogram based on multiscale products of eigen values of Hessian matrix. The detection of microcalcifications is achieved by decomposing the mammograms by filter bank based on Hessian matrix into different frequency sub-bands, suppressing the low-frequency subband, and finally reconstructing the subbands containing only significant high frequencies features. The significant features are obtained by multiscale products. Preliminary results indicate that the proposed scheme is better in suppressing the background and detecting the microcalcification clusters than any other detection methods.

Mehul P. Sampat et .al.,[13] in 2008 developed an algorithm for enhancement of spicules of spiculated masses, which uses the Discrete Radon Transform (DRT). Earlier they employed a commonly used method to compute the Discrete Radon Transform, which they refer to as the DRT. A new more exact method was developed by Averbuch et. al. to compute the Discrete Radon Transform which is called the Fast Slant Stack (FSS) method. The hypothesis was that this new formulation would help to improve their enhancement algorithm. To test this idea, they conducted multiple two-alternative-forced-choice observer studies and found that most observers preferred the enhanced images generated with the FSS method.

Heiko Hoffmann [14] in 2006 introduced and investigated the use of kernel principal component analysis (PCA) for novelty detection. Kernel PCA is a non-linear extension of PCA. Training data are mapped into an infinite-dimensional feature space. In this space, kernel PCA extracts the principal components of the data distribution. The squared distance to the corresponding principal subspace is the measure for novelty. This new method demonstrated a competitive performance on two-dimensional synthetic distributions and on two real-world data sets: handwritten digits and breast-cancer cytology.

_______ _______

Gilles Blanchard, [15] et. al., in 2004 investigated the effect of Kernel Principal Component Analysis (KPCA) within the classification framework, essentially the regularization properties of this dimensionality reduction method. KPCA has been previously used as a pre-processing step before applying an SVM but they point out that this method is somewhat redundant from a regularization point of view and they proposed a new algorithm called Kernel Projection Machine to avoid this redundancy, based on an analogy with the statistical framework of regression for a Gaussian white noise model. Preliminary experimental results show that this algorithm reaches the same performances as an SVM.

N.Toth, [16] in 2006, combined two independently developed methods for detecting cancer-indicating masses in mammograms. The first method is based on pixel intensities, the second one on texture features. The possible combinations of the two different approaches are investigated to achieve better mass detection rate with less false warnings. The composite system was tested with 523 mammographic cases, each containing 4 images.

C. Munteanu et .al.,[17] in 2006 introduced an interactive enhancement and optical perception evaluation tool that emphasizes breast masses from surrounding tissue in a digital mammogram and that is well suited for difficult cases featuring a high degree of subtlety for detecting the true masses. The tool is based on the Interactive Evolution of a Density Weighted Contrast Enhancement (DWCE) filter. After an average of 25 interactive enhancements of a given digital mammography, the physician is able to obtain a good enhancement (according to his/her particular criteria of analysis) of suspicious formations in the breast tissue. The tool produces enhanced images that allow for a further better discrimination between malign and benign masses, than if the original digital mammography is to be analyzed directly. Their results are presented using digital mammograms from the Digital Database for Screening Mammography (DDSM) available from the University of South Florida (USF).

Dave Tahmoush [18] in 2009 worked on image feature clustering to reduce the noise and the feature space, and the results are used in a distance function that uses a learned threshold in order to produce a classification. The threshold parameter of the distance function is learned simultaneously with the underlying clustering and then integrated to produce an agglomeration that is relevant to the images. He developed a method for differencing and classifying images which has incorporated into CAD. He also showed that using the image comparisons to determine the classification is insensitive to the parameters of the approach. He created and compared multiple models, demonstrating improved results over both academic and commercial approaches. He also defined a new distance measure for the comparison of point sets and demonstrates its effectiveness. The coupling of this distance measure with the parametric learning of clusters led to a highly effective classification technique.

S. Saheb Basha, [19] in 2009 presented a novel approach to identify the presence of breast cancer mass in mammograms. Their work utilizes morphological operators for segmentation and fuzzy c- means clustering for clear identification of clusters. The morphological operations and FCM is a new approach, using this they have successfully detected the breast cancer masses in mammograms.

Mohiy Hadhoud, [20] in 2005, introduced an algorithm for efficient detection of breast cancer tumor from digital mammograms images based on mathematical morphology and wavelet analysis. The algorithm contains two stages: enhancement and segmentation. Diagnosing cancer tissues using X-ray mammograms is a time consuming task even for highly skilled radiologists because mammograms are low contrast and noisy images. This assures the need for enhancement imaging to aid interpretation. For this purpose, they adopted mathematical morphology and wavelet-based-level dependent thresholding algorithms to increase the contrast in mammograms. Then, they adopted wavelet analysis to estimate a multiscale global thresholding in the segmentation process to extract suspicious regions known as regions of interest. Experimental results showed that the proposed algorithm by them yields significantly superior image quality when it is compared to the other well-known algorithms.

Arnau Oliver,, [21] in 2007 quantitatively compared the use of Fuzzy C-Means, Normalised Cuts and Mean Shift for grouping the tissue with similar grey-level, as a first step of a full strategy for classifying the breasts according their internal density They have found that, due to the internal nature of the algorithm (the use of the distance between pixels into the affinity function), the performance of the Normalised Cuts is lower compared to the use of both other algorithms. On the other hand, Mean Shift and Fuzzy C-Means obtained similar performance, around 78% correct classification. They considered these results in-line with the expected results, because when comparing the agreement between each individual expert annotations and the consensus opinion they obtained 78%, 89%, and 72% agreement. In addition, they effectively tested the use of Local Binary Patterns and co-occurrence matrices for describing the breast tissue textural information. The results obtained from the complete MIAS database and using a leave-one-woman-out strategy show that LBP and co-occurrence matrices features provide similar overall results, although LBP performs better in dense breasts while co-occurrence matrices in fatty breasts.

Li et al. [22] in 1995 developed a two-step for detection of masses. In the first step adaptive gray level thresholding was used to obtain an initial segmentation of suspicious regions. The segmentation was iteratively improved using a Multi-resolution markov Random Field (MRF) based segmentation method. The algorithm was first applied at the coarsest resolution and the output was refined at the next finer resolution. This strategy helps to reduce the computational complexity. In the second stage a fuzzy binary decision tree was used to classify the segmented regions as masses or normal tissue using features based on shape, region size and contrast.

Matsubara et al. [23] in 1996 developed an adaptive thresholding technique for the detection of masses. They employed histogram analysis techniques to divide mammograms into 3 categories ranging from fatty to dense tissue. Potential masses were detected using multiple threshold values based on the category of the mammogram. A number of features such as circularity, area, and standard deviation were used to reduce the number of false positives.

Li et al. [24] in 2001 developed a method for lesion site selection using morphological enhancement and stochastic model-based segmentation technique. A finite generalized Gaussian mixture distribution was used to model histograms of mammograms. The Expectation Maximization algorithm [25] was used to determine the parameters of the model. The segmentation was achieved by classifying pixels using a new Bayesian relaxation labeling technique. An underlying motivation for this technique was that it could incorporate neighborhood information into the classification process and that this would help improve the process. They argued that for the purpose of lesion site selection, sensitivity should be the sole criterion for evaluation and thus did not incorporate a false positive detection step.

R. Zwiggelaar et al., [26] in 1998 concentrated his work on the detection of spiculated lesions in mammograms. A spiculated lesion is typically characterized by an abnormal pattern of linear structures and a central mass. They have developed Statistical models to describe and detect both these aspects of spiculated lesions. They described a generic method of representing patterns of linear structures, which relies on the use of factor analysis to separate the systematic and random aspects of a class of patterns. They modeled the appearance of central masses using local scale-orientation signatures based on recursive median filtering, approximated using principal component analysis. For lesions of 16 mm and larger the pattern detection technique resulted in a sensitivity of 80% at 0.014 false positives per image, whilst the mass detection approach resulted in a sensitivity 80% at 0.23 false positives per image Simple combination techniques resulted in an improved sensitivity and specificity close to that required to improve the performance of a radiologist in a prompting environment.

Tetsuko ICHIKAWA et al., [27] in 2004 developed an automatic method for detecting areas of architectural distortion with spiculation. The suspect areas are detected by concentration indexes of line-structures extracted by using mean curvature. After that, discrimination analysis of nine features is employed for the classifications of true and false positives. The employed features are the size, the mean pixel value, the mean concentration index, the mean isotropic index, the contrast, and four other features based on the power spectrum. As a result of this work, the accuracy of the classification was 76% and the sensitivity was 80% with 0.9 false positives per image in our database in regard to spiculation. It was concluded that their method was effective in detecting the area of architectural distortion; however, some architectural distortions were not detected accurately because of the size, the density, or the different appearance of the distorted areas.

Rolando R. Hernandez-Cisneros et al., [28] in 2006 proposed a procedure for the classification of microcalcification clusters in mammograms using sequential Difference of Gaussian filters (DoG) and three Evolutionary Artificial Neural Networks (EANNs) compared against a feedforward Neural Network (NN) trained with backpropagation. They found that the use of Genetic Algorithms (GAs) for: finding the optimal weight set for a NN ; finding an adequate initial weight set before starting a backpropagation training algorithm and designing its architecture and tuning its parameters, results mainly in improvements in overall accuracy, sensitivity and specificity of a NN, compared with other networks trained with simple backpropagation.

Rangaraj M. Rangayyan, et al., [29] in 2010 investigated the effect of pixel resolution on texture features computed using the Gray-Level Co-Occurrence Matrix (GLCM) and was analyzed in the task of discriminating mammographic breast lesions as benign masses or malignant tumors. Regions in mammograms related to 111 breast masses, including 65 benign masses and 46 malignant tumors, were analyzed at pixel sizes of 50, 100, 200, 400, 600, 800, and 1,000 μm. Classification experiments using each texture feature individually provided accuracy, in terms of the Area Under the receiver operating characteristics Curve (AUC), of up to 0.72. Using the Bayesian classifier and the leave-oneout method, the AUC obtained was in the range 0.73 to 0.75 for the pixel resolutions of 200 to 800 μm, with 14 GLCM-based texture features using adaptive ribbons of pixels around the boundaries of the masses. Texture features computed using the ribbons resulted in higher classification accuracy than the same features computed using the corresponding regions within the mass boundaries. The texture features computed with the pixel size of 200 μm provided the highest average AUC with statistically highly significant differences as compared to all of the other pixel sizes tested, except 100 μm.

Roberto R. Pereira Jr. et al.,[30] in 2007 presents the usefulness of texture features in the classification of breast lesions in 5518 images of regions of interest, which were obtained from the Digital Database for Screening Mammography(DDSM) that included microcalcifications, masses, and normal cases. Sixteen texture features were used, i.e., 13 were based on the spatial gray-level dependence matrix and 3 on the wavelet transform. The nonparametric K-NN classifier was used in the classification stage. The results obtained from Receiver Operating Characteristics (ROC) analysis indicated that the texture features can be used for separating normal regions and lesions with masses and microcalcifications, yielding the Area Under the Curve (AUC) values of 0.957 and 0.859, respectively. The study showed that the texture features can be used for the detection of suspicious regions in mammograms.

Lubomir Hadjiiski et al., [31] in 2008 developed an automated system for detecting microcalcifications within a predefined region of interest (ROI), and classifying the clusters as malignant and benign on Full-Filled Digital Mammograms (FFDM). Their system consists of two stages. In the first stage, a detection program is used to detect cluster candidates within the ROI. A rule-based identification method is designed to differentiate the true and false clusters. In the second stage, morphological and texture features are extracted from the selected clusters and a classifier is trained to classify malignant and benign clusters. In this study, they used a data set of 247 ROIs (63 malignant and 184 benign) containing biopsy-proven calcification clusters were used. An MQSA radiologist identified 117 corresponding clusters on the CC and MLO pairs of mammograms. Leave-one-case-out resampling was used for feature selection and classification. Two MQSA radiologists evaluated the two view pairs. The detection program correctly detected 100% (247/247) of the clusters of interest with 0.14 (35/247) FPs/ROI. The identification program correctly selected 99.2% (245/247) of the index clusters. In the classification stage an average of 4 features was selected from the training subsets. The most frequently selected features included 3 morphological and 1 texture features. The classifier achieved a test Az of 0.73 for classifying the 247 clusters as malignant or benign. For the 117 pairs of matched CC and MLO views the test Az was 0.77. The partial area index above a sensitivity of 0.9, Az(0.9) was 0.21. In comparison, the two experienced MQSA radiologists achieved Az of 0.76 and 0.73, respectively, for the 117 CC and MLO view pairs. The partial area index Az (0.9) was 0.27 and 0.12, respectively. Their classification system can detect the microcalcifications within the specified ROI on mammogram with high sensitivity and satisfactory specificity, and classified them with accuracy comparable to that of an experienced radiologist.

G. Ertas et al., [32] in 2001 investigated the use of shape features to classify breast masses and a classification scheme has been developed to classify masses as either benign or malignant. In this study, they calculated geometric parameters such as area, perimeter, circularity, normalized circularity, radial distance mean and standard deviation, area ratio, orientation, eccentricity, moment invariants and Fourier descriptors up to 10. A mammogram database designed to store the images of the masses, calculated shape descriptor parameters and some additional data, such as patient history, category of the mass and biopsy report if performed which are required in BI-RADS is also introduced A touch on memory system has been used as a tool that permits access to the electronic patient record in the mammogram database. The software was written in Delphi and runs on Windows operation systems.

Anna N. Karahaliou et al., [33] in 2006 conducted a study investigates whether texture properties of the tissue surrounding microcalcification (MC) clusters can contribute to breast cancer diagnosis. The case sample analyzed consists of 100 mammographic images, originating from the Digital Database for Screening Mammography (DDSM). All mammograms selected correspond to heterogeneously and extremely dense breast parenchyma and contain subtle MC clusters (46 benign and 54 malignant, according to database ground truth tables). Regions of interest (ROIs) of 128x128 pixels, containing the MCs are used for the subsequent texture analysis. ROIs are preprocessed using a wavelet-based locally adapted contrast enhancement method and a thresholding technique is applied to exclude MCs. Texture features are extracted from the remaining ROI area (surrounding tissue) employing first and second order statistics algorithms, grey level run length matrices and Laws' texture energy measures. Differentiation between malignant and benign MCs is performed using a k-nearest neighbour (kNN) classifier and employing the leave-one-out training-testing methodology. The Laws' texture energy measures demonstrated the highest performance achieving an overall accuracy of 89%, sensitivity 90.74% (49 of 54 malignant cases classified correctly) and specificity 86.96% (40 of the 46 benign cases classified correctly). Texture analysis of the tissue surrounding MCs shows promising results in computer-aided diagnosis of breast cancer and may contribute to the reduction of benign biopsies.

Anna N. Karahaliou et al., [34] in 2008 conducted a study to investigate texture properties of the tissue surrounding microcalcification (MC) clusters on mammograms for breast cancer diagnosis. The case sample analyzed consists of 85 dense mammographic images, originating from the Digital Database for Screening Mammography (DDSM). Mammograms analyzed contain 100 subtle MC clusters (46 benign and 54 malignant). The tissue surrounding MCs is defined on original and wavelet decomposed images, based on a redundant discrete wavelet transform. Gray-level texture and wavelet coefficient texture features at three decomposition levels are extracted from Surrounding Tissue Regions of Interest (ST-ROIs). Specifically, gray-level first-order statistics, gray-level cooccurrence matrices features, and Laws' texture energy measures are extracted from original image ST-ROIs. Wavelet coefficient first-order statistics and wavelet coefficient cooccurrence matrices features are extracted from subimages ST-ROIs. The ability of each feature set in differentiating malignant from benign tissue is investigated using a probabilistic neural network. Classification outputs of most discriminating feature sets are combined using a majority voting rule. The proposed combined scheme achieved an area under receiver operating characteristic curve (Az) of 0.989. Results suggested that MCs surrounding tissue texture analysis can contribute to computer-aided diagnosis of breast cancer.

Pelin Gorgel et al., [35] in 2009 conducted a study to investigate an approach for classification of mammographic masses as benign or malign. The study relies on a combination of Support Vector Machine (SVM) and wavelet-based subband image decomposition. Decision making was performed in two stages as feature extraction by computing the wavelet coefficients and classification using the classifier trained on the extracted features. SVM, a learning machine based on statistical learning theory, was trained through supervised learning to classify masses. The research involved 66 digitized mammographic images. The masses were segmented manually by radiologists, prior to introduction to the classification system. Preliminary test on mammogram showed over 84.8% classification accuracy by using the SVM with Radial Basis Function (RBF) kernel. Also confusion matrix, accuracy, sensitivity and specificity analysis with different kernel types were used to show the classification performance of SVM.

Rabi Narayan Panda et al., [36] in 2009 proposed a research intended to develop an image processing algorithm for the recognition of microcalcifications and mass lesions to aid the premature detection of breast cancer. The work proposed deals with a novel approach for the extraction of features like microcalcifications and mass lesions in mammograms for early detection of breast cancer. The proposed technique is based on a three-step procedure: (a) regions of interest (ROI) specification, (b) two dimensional wavelet transformation, and (c) feature extraction based on OTSU thresholding the region of interest for the identification of microcalcifications and mass lesions. ROIs are preprocessed using a wavelet-based transformation method and a thresholding technique is applied to exclude microcalcifications and mass lesions. The method suggested for the detection of microcalcifications and mass lesions from mammogram image segmentation and analysis was tested over several images taken from mini-MIAS (Mammogram Image Analysis Society, UK) database. The implementation of the algorithm was carried out using Matlab codes programming and thus is capable of executing effectively on a simple personal computer with digital mammogram as accumulated data for assessment.

Michael A. Yacoub et al.,[37] in 2006 proposed a method to detect malignant tumors. The method is a three-step process. The first step is ROI extraction of 256 x 256 pixels size. The second step is the feature extraction, where we used a set of 99 features and we found that 83 of these feature are capable of differentiating between normal and cancerous breast tissues. The third step is the classification process. We used the techniques of the minimum distance, the k-Nearest Neighbor (k-NN) and Bayes classifiers to classify between normal and cancerous tissues. We examined the effect of changing the size of ROI extracted from the mammogram on the system by extracting ROI of size 512 x 512. Their computerized scheme was shown to have the potential to detect malignant tumors with a clinically acceptable sensitivity and low false positives.

Yu Zhang et al, [38] in 2009 presented a novel segmentation method for identifying mass regions in mammograms. The aim of the work is to build a Computer-Aided Diagnosis (CADx) system that classifies suspicious cancer masses in mammograms as benign or malignant. Segmentation of suspicious mass regions is an important pre-processing step to achieve high accuracy results, because the regions of interest (ROI) marked by radiologists are oftentimes imprecise. For each ROI, they first applied an enhancement function to increase the grey-level contrast of the image, then a filter function to reduce noise. Next, for each pixel in the ROI, they computed the energy feature based on the co-occurrence matrix of the pixel. Finally they extracted a contour of the mass from the energy feature image using an edge-based segmentation technique.

Maurice Samulski et al,[39] in 2007 compared two state-of-the-art classification techniques characterizing masses as either benign or malignant, using a dataset consisting of 271 cases (131 benign and 140 malignant), containing both a MLO and CC view. For suspect regions in a digitized mammogram, 12 out of 81 calculated image features have been selected for investigating the classification accuracy of support vector machines (SVMs) and Bayesian networks (BNs). Additional techniques for improving their performance were included in their comparison: the Manly transformation for achieving a normal distribution of image features and Principal Component Analysis (PCA) for reducing our high-dimensional data. The performance of the classifiers was evaluated with Receiver Operating Characteristics (ROC) analysis. The classifiers were trained and tested using a k-fold cross-validation test method (k=10). It was found that the area under the ROC curve (Az) of the BN increased significantly (p=0.0002) using the Manly transformation, from Az = 0.767 to Az = 0.795. The Manly transformation did not result in a significant change for SVMs. Also the difference between SVMs and BNs using the transformed dataset was not statistically significant (p=0.78). Applying PCA resulted in an improvement in classification accuracy of the naive Bayesian classifier, from Az = 0.767 to Az = 0.786. The difference in classification performance between BNs and SVMs after applying PCA was small and not statistically significant (p=0.11).

Indra Kanta Maitra et al.[40], in 2011 developed a method to make a supporting tool for Identification of Abnormal Masses in Digital Mammography Images. This will make it easy and less time consuming for identification of abnormal masses in digital mammography images. The mammogram images used in the experiment are taken from the mini mammography database of MIAS. The original MIAS Database (digitized at 50 micron pixel edge) has been reduced to 200-micron pixel edge and clipped/padded so that every image is 1024 pixels x 1024 pixels. All images are held as 8-bit gray level scale images with 256 different gray levels (0-255) and physically in portable gray map (.pgm) format. The identification technique is divided into two distinct parts i.e. Formation of Homogeneous Blocks and Color Quantization after preprocessing. In their experiment they have considered all three types of breast tissues i.e. Fatty, Fatty-glandular, Dense-glandular. Different types of abnormalities like calcification, well-defined or circumscribed masses, speculated masses and other ill-defined masses have also been considered for their experimentation.

Tiffany Tweed et al., [41] in 2002 presented an algorithm that selects regions of interest (ROI) containing a tumour in mammogram, based on the combination of a texture and histogram analysis. The first analysis compares texture features extracted from different regions in an image to the same features extracted from known tumorous regions. The second analysis detects the ROI with two thresholds computed from the histograms of known tumorous masks.

Nora Szekely et al, [42] in 2004 proposed a hybrid system for detecting masses in mammographic images. The proposed approach analyses the mammograms in three major steps. First a global segmentation method is applied to find the regions of interest. This step uses texture features, decision trees and a multiresolution Markov Random Field model. The second stage works on the output of the previous algorithm. Here a combination of three different local segmentation methods is used, and then some relevant features are extracted. Some of them refer to the shape of the object, others are simple texture parameters.

R. Nithya et al., [43] in 2011 evaluated and compared the performance of three different feature extraction methods for classification of normal and abnormal patterns in mammogram. Three different feature extraction methods used here are intensity histogram, GLCM (Grey Level Co-occurrence Matrix) and intensity based features. A supervised classifier system based on neural network is used. The performance of the each feature extraction method is evaluated on Digital Database for Screening Mammography (DDSM) breast cancer database. Their experimental results suggested that GLCM method outperformed the other two methods.

Sara Dehghani et al., [44] in 2011 proposed a method to improve preprocessing of Mammograms. The method has three phases. The first phase is omitting the excessive image parts which are in the two sides of the image; they did this work by the usage of the pixels brightness. The second phase is the distinction of the breast direction and put all images in one direction; they did this work by the usage of threshold limit of gray level of the two halves of the image. The third phase is the breast region segmentation from the background; they did this work by the usage of series of point operations and the growing region method and the result has been reported to 99%.

N. Riyahi Alam et al, [45] in 2009 developed a novel hybrid segmentation method has been developed for detection of masses in digitized mammograms using three parallel approaches: adaptive thresholding method, Gabor filtering and fuzzy entropy feature as a Computer-Aided Detection (CAD) scheme. The algorithm consists of the following steps: a) Preprocessing of the digitized mammograms including identification of region of interest (ROI) as candidate for massive lesion through breast region extraction, b) Image enhancement using linear transformation and subtracting enhanced from the original image, c) Characterization of the ROI by extracting the fuzzy entropy feature, d) Local adaptive thresholding for segmentation of mass areas, e) Filtering the input images using Gabor functions, f) Combine expert of the last three parallel approaches for mass detection. The proposed method was tested on 78 mammograms (30 normal & 48 cancerous) from the BIRADS and local databases. The detected regions validated by comparing them with the radiologists' hand-sketched boundaries of real masses. The current algorithm can achieve a sensitivity of 90.73% and specificity of 89.17%. This approach showed that the behavior of local adaptive thresholding, Gabor filters and fuzzy entropy technique could be useful for mass detection on digitized mammograms. The results suggest that their proposed method could help radiologists as a second reader in mammographic screening of masses.

Ying Cao et al., [46] in 2010, attempted to accurately segment the mammographic masses and distinguish malignant from benign tumors. An adaptive region growing algorithm with hybrid assessment function combined with maximum likelihood analysis and maximum gradient analysis was developed in their experiment. In order to accommodate different situations of masses, the likelihood and the edge gradients of segmented masses were weighted adaptively by the use of information entropy. 106 benign and 110 malignant tumors were included in the study. They found that the proposed algorithm obtained segmentation contour more accurately and delineated the tumor body as well as tumor peripheral regions covering typical mass boundaries and some spiculation patterns. Then the segmented results were evaluated by the classification accuracy. 42 features including age, intensity, shape and texture were extracted from each segmented mass and support vector machine (SVM) was used as a classifier. The classification accuracy was evaluated using the area (Az) under the receiver operating characteristic (ROC) curve. It was found that the maximum likelihood analysis achieved an Az value of 0.835, the maximum gradient analysis got an Az value of 0.932 and the hybrid assessment function performed the best classification result where the value of Az was 0.948. In addition, compared with traditional region growing algorithm, their proposed algorithm is more adaptive and provides a better performance for future works.

Aboul Ella et al., [47] in 2003 presented an efficient classification algorithm in digital mammograms in the context of rough set theory. Feature extractions acquired in the work are derived from the gray-level co-occurrence matrix. The features are extracted, normalized and then the rough set dependency rules are generated directly from the real value attribute vector. Then the classification is built and the quadratic distance function used to determines similarity between a query and database image. The experimental results show that the proposed algorithm performs well reaching over 85% in accuracy.

M. Hanmandlu et al., [48] in 2008 introduced two separate techniques for mass and micro-calcification segmentation in digital mammograms. Segmentation of masses consists of three steps- background subtraction, fuzzy texture representation and entropic thresholding. Similarly micro-calcifications are also segmented in three stages - background subtraction, Laplacian of Gaussian filtering and contrast estimation followed by thresholding. Both the techniques are verified with the markings given by the radiologist and are found to be quite effective tools in diagnosing breast cancer.

Leonardo de Oliveira Martins et al., [49] in 2009 presented a methodology for masses detection on digitized mammograms using the K-means algorithm for image segmentation and co-occurrence matrix to describe the texture of segmented structures. Classification of these structures is accomplished through Support Vector Machines, which separate them in two groups, using shape and texture descriptors: masses and non-masses. The methodology obtained 85% of accuracy.

Mahua Bhattacharya et al., [50] in 2009 attempted to segment the masses accurately and distinguish malignant from benign masses. The suspicious location of the breast masses are specified by the radiologists and then masses are accurately segmented using fuzzy c-means clustering technique. Fourier descriptors are utilized for the extraction of shape features of mammographic masses. These shape features along with the texture features are fed to the input of the Adaptive Neuro Fuzzy Network classifier for determination of the masses as benign, lobular or malignant. The classification system utilizes a simple Euclidian distance metric to determine the degree of malignancy. The study involves 40 digitized mammograms from MIAS, BIRADS database and has to be found 87% correct classification rate.

B. Senthilkumar et al., [51] in 2012 made improvements in region growing image segmentation for mammogram images to detect the breast cancer. Selective median filter is used for preprocessing, CLAHE (Contrast Limited Adaptive Histogram Equalization) method is used for the enhancement, Harris corner detect theory is used to auto find growing seeds and the seeded region growing rule for the development of regions. This work also includes a new uncertainty theory-Cloud Model to realize automatic and adaptive segmentation threshold selecting, which considers the uncertainty of image and extracts concepts from characteristics of the region to be segmented like human being. They found this method works reliable on homogeneity and region characteristics. Furthermore, the method has been tested for over 30 sample images and the results found were good.

Ali Cherif Chaabani et al., [52] in 2010 proposed a method of pre-processing on Medio-Lateral Oblique-view (MLO) mammograms that is composed of two stages: the first step helps to extract the breast region from the rest of the image (background.), while the second aims at the suppression of the pectoral muscle. To extract the breast region, they used a method based on automatic thresholding (Otsu's) and Connected Component Labeling algorithm. Identifying the pectoral muscle has been done using the Hough transform and active contour. They evaluated their pre-processing method on a set of 80 images obtained from the DDSM database and they found that breast region extraction gave an excellent success rate that reached 100%. The success rate in the removal of the pectoral muscle was 92.5% with the use of Hough transform and active contour.

Homero Schiabel et al., [53] in 2008 attempts to describe a methodology for segmenting suspect masses in dense breast images as a part of a CAD scheme under development. This methodology is based on the Watershed transformation, which is combined with two other procedures: a histogram equalization, working as pre-processing for enhance images contrast, and a labeling procedure intended to reduce noise. Tests with a set of 252 regions of interest extracted from 130 digitized mammograms have registered a scheme sensibility of 92% with about 90% of specificity. These results are promising when applied to dense breast images, which can improve significantly the performance of a processing scheme for such type of cases in mammography.

Jun Liu et al., [54] in 2010 proposed a new mass segmentation algorithm. In the proposed algorithm, a fully automatic marker-controlled watershed transform is proposed to segment the mass region roughly, and then a level set is used to refine the segmentation. For over-segmentation caused by watershed, they also investigated different noise reduction technologies. Images from DDSM were used in the experiments and the results show that the new algorithm can improve the accuracy of mass segmentation. The combination of the watershed based segmentation and level set method can improve the efficiency of the segmentation.

Fatima Eddaoudi et al., [55] in 2011 carried their work on focusing the masses detection using SVM classification and textures analysis. The identification of tumors is generally done in three stages: pectoral muscle segmentation, hard density zone detection and texture analysis of regions of interest. As for the first stage, pectoral muscle they used an approach based on contour detection using snakes with an automatic initialization. For the second stage, they used an approach based on maxima thresholding. The region of interesting segmented are classified to normal and abnormal tissue using Haralick features calculated from the cooccurrence matrix. The test of these methods on mammograms of MIAS databases showed better performance in detecting masses compared to the methods proposed in the literature.

Joan Marti et al., [56] in 2003 proposed a supervised method for the segmentation of masses in mammographic images. The algorithm starts with a selected pixel inside the mass, which has been manually selected by an expert radiologist. Based on the active region approach, an energy function is defined which integrates texture, contour and shape information. Then, pixels are aggregated or eliminated to the region by optimizing this function allowing to obtain an accurate segmentation. Moreover, a texture feature selection process, performed before the segmentation, ensures a reliable subset of features. Experimental results prove the validity of the proposed method.

Wael A. Mohamed et al., [57] in 2007 introduced Computer Aided Diagnosis system for Digital Mammograms, as an aid to radiologists, a computer diagnosis system, which could be helpful in diagnosing abnormalities faster than traditional screening program without the drawback attribute to human factors. The techniques used for feature extraction is based on the invariant features and fractal dimensions of locally processed image (ROI). Two statistical classifiers (The minimum distance classifier and the voting K-Nearest Neighbor classifier) were used and compared through the system to reach a better classification decision.

Jawad Nagi et al., [58] in 2010 proposed an algorithm for Breast profile segmentation for ROI Detection in Digital Mammograms. Finding an accurate, robust and efficient breast profile segmentation technique still remains a challenging problem in digital mammography. Extraction of the breast profile region and the pectoral muscle is an essential pre-processing step in the process of computer-aided detection. Primarily it allows the search for abnormalities to be limited to the region of the breast tissue without undue influence from the background of the mammogram. The presence of pectoral muscle in mammograms biases detection procedures, which recommends removing the pectoral muscle during mammogram pre-processing. They explored an automated technique for mammogram segmentation. The proposed algorithm uses morphological preprocessing and seeded region growing (SRG) algorithm in order to: (1) remove digitization noises, (2) suppress radiopaque artifacts, (3) separate background region from the breast profile region, and (4) remove the pectoral muscle, for accentuating the breast profile region. To demonstrate the capability of their proposed approach, digital mammograms from two separate sources are tested using Ground Truth (GT) images for evaluation of performance characteristics. Experimental results obtained indicate that the breast regions extracted accurately correspond to the respective GT images.

Ibrahima Faye et al., [59] in 2009, introduced a new method of feature extraction from Wavelet coefficients for classification of digital mammograms. A matrix is constructed by putting Wavelet coefficients of each image of a building set as a row vector. The method consists then on selecting by threshold, the columns which will maximize the Euclidian distances between the different class representatives. The selected columns are then used as features for classification. The method is tested using a set of images provided by the Mammographic Image Analysis Society (MIAS) to classify between normal and abnormal and then between benign and malignant tissues. For both classifications, a high accuracy rate (98%) is achieved by them.

Mazen E. Osman et al [60] in 2009 proposed a method to develop a Computer-Aided Diagnostic system for classification of microcalcifications in digital mammograms; it splits into three-step process. The first step is Region of Interest extraction of 32 x 32 pixels size. The second step is the features extraction, where we used a set of 234 features from Region of Interest by employing wavelet decomposition, 1st order statistics from wavelet coefficients algorithms; also, they extracted 1st order statistics, median contrast and local binary partition features. The third step is the classification process where differentiation between normal and abnormal is performed using a Minimum Distance Classifier and K-Nearest Neighbor Classifiers employing the leave-one-out training-testing methodology. The results show acceptable sensitivity and specificity for the proposed system.

H.B.Kekre et al., [61] in 2010 devised a new algorithm for texture based segmentation using statistical properties. For that probability of each intensity value of image is calculated directly and image is formed by replacing intensity by its probability. Variance is calculated in three different ways to extract the texture features of the mammographic images. These results of proposed algorithm are compared with well known GLCM and Watershed algorithm.

R. Nithya et al., [62] in 2011 proposed a method for breast cancer diagnosis in digital mammograms using GLCM (Grey Level Co-occurrence Matrix) features. They developed Computer Aided Diagnosis (CAD) system using GLCM feature and neural network. They have extracted five GLCM features for each mammogram image. Mammogram image is classified into normal image and cancer image. The effectiveness of the proposed method has examined on DDSM (Digital Database for Screening Mammography) database using classification accuracy, sensitivity and specificity. The overall accuracy can be improved by most relevant GLCM features, which is selected by feature selection algorithm.

Erkang Cheng et al.,[63] in 2010 proposed Histogram Intersection(HI) method for mammographic image classification. First, they used the bag-of-words framework for image representation, which captures the texture information by collecting local patch statistics. In this framework, they first represented an image by treating it as a set of local patches. These patches are then vector quantized according to a codebook learned from training images. Then an image is represented by the code histogram of its patches. Such histograms are then compared for classification. Then, they proposed using normalized histogram intersection (HI) as a similarity measure with the K-nearest neighbor (KNN) classifier. Furthermore, by taking advantage of the fact that HI forms a Mercer kernel, they combined HI with support vector machines (SVM), which further improves the classification performance. The proposed methods are evaluated on a galactographic dataset and are compared with several previously used methods. In a thorough evaluation containing about 288 different experimental configurations, the proposed methods demonstrate promising results.

M. Wirth et al., [64] in 2005 explored a new algorithm for breast region segmentation using fuzzy reasoning. The algorithm uses morphological pre-processing to suppress artifacts and accentuate the breast region, followed by a fuzzy rule-based algorithm to classify the breast tissue region. To demonstrate the capability of our segmentation algorithm it is extensively tested on mammograms from two databases using ground truth images and quantitative metrics to evaluate its performance characteristics. The experimental results indicate that the breast regions extracted accurately characterize the corresponding ground truth images. The algorithm is fully autonomous, and is able to preserve the skin and nipple (if in profile), a task very few existing mammogram segmentation algorithms can claim.

Issam El-Naqa et al., [65] in 2002 investigated an approach based on Support Vector Machines (SVMs) for detection of microcalcification (MC) clusters in digital mammograms, and propose a successive enhancement learning scheme for improved performance. They formulated MC detection as a supervised-learning problem and apply SVM to develop the detection algorithm. They used the SVM to detect at each location in the image whether an MC is present or not. They tested the proposed method using a database of 76 clinical mammograms containing 1120 MCs. They used Free-Response Receiver Operating Characteristic (FROC) curves to evaluate detection performance, and compare the proposed algorithm with several existing methods. In their experiments, the proposed SVM framework outperformed all the other methods tested. In particular, a sensitivity as high as 94% was achieved by the SVM method at an error rate of one false-positive cluster per image. The ability of SVM to outperform several well-known methods developed for the widely studied problem of MC detection suggests that SVM is a promising technique for object detection in a medical imaging application.

H. Mirzaalian et al.,[66] in 2007 proposed one low-pass mask for detecting breast contour and a new method for the identification of the pectoral muscle in most Medio-Lateral Oblique(MLO) mammograms based on Non-Linear Diffusion algorithm which is an edge preserving smoother. Evaluation of the breast contour and pectoral muscle detected in the mammograms were performed by the Hausdorff Distance Measure (HDM) and also the Mean of Absolute Error Distance Measure (MAEDM) based on a distance transform and image algebra between the edges identified by radiologists and by the proposed methods. Then they compared their results by other segmentation methods. Their proposed algorithms show superior results in comparison.

E. A. Zanaty et al., [67] in 2008 introduced a new kernel function that could improve the SVMs classification accuracy. The proposed kernel function, called Polynomial Radial Basis Function (PRBF) combines both Gauss (RBF) and Polynomial (POLY) kernels. They proved that the proposed kernel converges faster than the Gauss and Polynomial kernels and also gives a good classification accuracy in nearly all the data sets specially with high dimension ones. Thereafter SVMs algorithm based on the PRBF is implemented and experimented with non-separable data set with several attributes to prove its efficiency. Then, the obtained results are compared with SVMs algorithms that are based on Gaussian and Polynomial kernels.

Anthony Nguyen et al.,[68] in 2009 identified several aspects of the image characteristics relevant to viewer perception, including intensity properties (such as contrast), spatial properties (such as texture) and structure properties (such as breast density). They provided quantitative descriptions of the variability of these properties over a test set of 12 screening mammograms drawn from three different modalities and containing a typical mix of screening cases.

Y.Ireaneus Anna Rejani et al.,[69] in 2009 presented a tumor detection algorithm from mammogram. The system focused on the solution of two problems. One is how to detect tumors as suspicious regions with a very weak contrast to their background and another is how to extract features which categorize tumors. The tumor detection method follows the scheme of (a) mammogram enhancement. (b) The segmentation of the tumor area. (c) The extraction of features from the segmented tumor area. (d) The use of SVM classifier.

The enhancement can be defined as conversion of the image quality to a better and more understandable level. The mammogram enhancement procedure includes filtering, top hat operation, DWT. Then the contrast stretching is used to increase the contrast of the image. The segmentation of mammogram images has been playing an important role to improve the detection and diagnosis of breast cancer. The most common segmentation method used is thresholding. The features are extracted from the segmented breast area. Next stage includes, which classifies the regions using the SVM classifier. The method was tested on 75 mammographic images, from the mini-MIAS database. The methodology achieved a sensitivity of 88.75%.

Viet Dzung Nguyen et al., [70] in 2009 developed an automated CAD system for detection and classification of massive lesions in mammographic images is presented. The system consists of three processing steps: Regions-Of- Interest detection, feature extraction and classification. Our CAD system was evaluated on Mini-MIAS database consisting 322 digitalized mammograms. The CAD system's performance is evaluated using Receiver Operating Characteristics (ROC) and Free Response ROC (FROC) curves. They achieved results of 3.47 False Positives per Image (FPpI) and sensitivity of 85%.

[1] David J. Marchette, Richard A. Lorey Naval Surface Warfare Center, B10, Dahlgren, VA 22448 Carey E. Priebe Johns Hopkins University, Dept. of Math. Sciences, Baltimore, MD "An Analysis of Local Feature Extraction in Digital Mammography" 21218.

[2]Dimitris K. Iakovidis, Member, IEEE, Nikos Pelekis, Haralampos Karanikas, Evangelos Kotsifakos, Ioannis Kopanakis, and Yannis Theodoridis "A Pattern Similarity Scheme for Medical Image Retrieval" , 2006.

[3]H.S.Sheshadri, A Kandaswamy, "Computer Aided Diagnosis of Digital Mammograms" Information Technology Journal 5 (2) pp.342-346 ISSN 1812-5638, 2006.

[4] Nuryanti Mohd. Salleh, Harsa Amylia Mat Sakim and Nor Hayati Othman, "Neural Networks to Evaluate Morphological Features for Breast Cells Classification" International Journal of Computer Science and Network Security, Vol.8 No.9, September 2008.

[5]Matteo Roffilli, "Advanced Machine Learning Techniques for Digital Mammography", Technical Report UBLCS-2006-12 March 2006.

[6] Akram I. Omara, Ahmed S. Mohamed, Abo-Bakr M. Youssef, and Yasser M. Kadah "Computer Aided Diagnosis in Digital Mammography" Proc. of Cairo International Biomedical Engineering Conference 2006.

[7] Yih-Chih Chiou, Cheng-Yu Lind an Yu-Teng Liang, "A New Method for the Registration of Mammograms" JCIS-2006 Proc. of Advances in Intelligent Systems Research ISBN 978-90-78677-01-7 , October 2006.

[8] Michal Haindl and Stanislav Mike, "Unsupervised Mammograms Segmentation" 19th International Conference on Pattern Recognition, ISSN: 1051-4651, 2008.

[9] K.Thangavel and M.Karnan and R.Sivakumar and A. Kaja Mohideen, "Ant Colony System for Segmentation and Classification of Microcalcification in Mammograms" The International Journal of Artificial Intelligence and Machine Learning Volume 5, Issue III, pp. 29-40, 2005.

[10] Arianna Mencattini, Giulia Rabottino, Marcello Salmeri, Federica Caselli, Roberto Lojacono and Giulia Frondizi, Features Extraction to Classify Microcalcification Clusters in Mammography, 16th IMEKO TC4 Symposium on Exploring New Frontiers of Instrumentation and Methods for Electrical and Electronic Measurements, Sept. 2008.

[11] Sheshadri HS, and Kandaswamy A "Detection of breast cancer by mammogram image segmentation" Journal of Cancer Research Therapeutics (JCRT) - Volume 1 - Issue 4, December 2005.

[12] T.Balakumaran and Dr. Ila Vennila, "A Computer Aided Diagnosis System for Microcalcification Cluster Detection in Digital Mammogram", International Journal of Computer Applications (0975 - 8887) pp. 39-45, Volume 34, No.1, November 2011.

[13] Mehul P. Sampat, Gary J. Whitman, Alan C. Bovik and Mia K. Markey, "Comparison of Algorithms to Enhance Spicules of Spiculated Masses on Mammography", Journal of Digital Imaging, Volume 21, No. 1, pp 9-17, March 2008.

[14] Heiko Hoffmann, Kernel PCA for Novelty Detection, Pattern Recognition 2006.

[15] Gilles Blanchard, Pascal Massart, Regis Vert, and Laurent Zwald, "Kernel Projection Machine: a New Tool for Pattern Recognition", In the Proc. of Advances in Neural Information Processing Systems [NIPS] December 13-18, 2004.

[16] N. Toth, G. Takacs and B. Pataki, "Mass Detection In Mammograms Combining Two Methods", Department of Measurement and Information Systems, Budapest University of Technology and Economics, 2006.

[17] C. Munteanu, E. M. Suarez, M. A. Rodrıguez-Florido, R. Cardenes and J. Ruiz-Alzola , "Emphasizing Breast Masses in Digital Mammography by Interactive Evolution of a Density Weighted Contrast Enhancement Filter", Proc. of the IEEE-EMB International Special Topic Conference on Information Technology in Biomedicine ,2006

[18] Dave Tahmoush, "Image Similarity to Improve the Classification of Breast Cancer Images", Open-Access article: algorithms ISSN 1999-4893, 2009.

[19] S.Saheb Basha and K.Satya Prasad, "Automatic Detection of Breast Cancer Mass in

Mammograms us