Introducing kernel based morphology as an enhancement method for mass classification on mammography.

Since mammography images are in low-contrast, applying enhancement techniques as a pre-processing step are wisely recommended in the classification of the abnormal lesions into benign or malignant. A new kind of structural enhancement is proposed by morphological operator, which introduces an optimal Gaussian Kernel primitive, the kernel parameters are optimized the use of Genetic Algorithm. We also take the advantages of optical density (OD) images to promote the diagnosis rate. The proposed enhancement method is applied on both the gray level (GL) images and their OD values respectively, as a result morphological patterns get bolder on GL images; then, local binary patterns are extracted from this kind of images. Applying the enhancement method on OD images causes more differences between the values therefore a threshold method is applied toremove some background pixels. Those pixels that are more eligible to be mass are remained, and some statistical texture features are extracted from their equivalent GL images. Support vector machine is used for both approaches and the final decision is made by combining these two classifiers. The classification performance rate is evaluated by Az, under the receiver operating characteristic curve. The designed method yields Az = 0.9231, which demonstrates good results.

INTRODUCTION

Breast cancer excludes lung and bronchus cancer has a high mortality rate among woman as American Cancer society reported.[1] The reason is not recognized yet, but early breast cancer detection can be effective in patients’ recovery. Despite medical image improvement such as magnetic resonance imaging and ultrasound, mammography is still one of the effective solutions for the detection of breast cancer, and reading mammography images is a demanding job for radiologists. Therefore, computer-aided design (CAD) systems are used to help radiologists in mammography images analysis in the case of breast abnormality detection such as mass, calcification, and architectural distortion.

Since lesion biopsy came to benign in 65-90% of cases, development of the CAD systems is very important in classifying the lesion to malignant and benign in order to reduce unnecessary biopsy. The diagnosis accuracy is increased by merging CAD systems with radiologist knowledge.[2] The CAD effectiveness depends on the accuracy rate of lesion classification and the extracted feature from the mammography images. Mass detection and classification are more challenging tasks than the other abnormalities because masses are not distinguished from the surrounding tissues. Cheng et al.[3] presented a review for mass detection and classification methods.

As the contrast of mammography images is mostly low, image enhancement techniques are often used to enhance the images that increase radiologist diagnosis performance and help the CAD systems. An enhancement image is an image in which the observer takes a better perception for the desirable information. In the case of abnormal tissues, structural elements differ from benign masses to malignant ones. Enhancing these structures can cause a better diagnosis rate on mammography images. Since, the Morphological operations are used for different purposes of image processing, mainly image enhancement,[45] we introduce a kind of morphological operator in which the primitive structure differs from the past.

The purpose of our method is to enhance suspected lesions so that the early stage diagnosis can be carried out in an effective manner. The new method causes the structural elements to be more visible with the help of morphology operation in which the primitive is estimated by the Gaussian kernel. Local binary pattern (LBP) is extracted from enhanced images, which is a kind of structural texture feature and defines the spatial structure of an image. LBP has been widely used in face recognition[678] and other applications such as cancer detection. In a related work in 2011, Jun Lui et al. used an improved LBP in mass classification.[9] The proposed approach also takes advantages from optical density (OD) images,[10] which are free from scanner type and are a measure of blackness[11] that distinguish small differences.[12] In this paper, using this kind of images is shown to be effective and beneficial in mass classification.

The paper is organized as follows. In section 2, necessary pre-processing methods for the classification task are cited. Section 3 describes the proposed approach, and section 4 and 5 include the results and discussion respectively.

MATERIALS AND METHODS

Mammograms Dataset

The mammography cases are from digital database for screening mammography (DDSM), which is publicly available by University of South Florida.[13] The selected cases are taken by different scanners, but they are all digitalized with 12 bit depth and 50 μm pixel size. Some of the images are omitted because of some artifacts such as pencil mark, clips shadow, masses with calcifications, and other abnormalities. The new dataset consists of 516 regions (340 benign and 176 cancers) from 345 patients.

Region of Interest Extraction

The ROI is the same as estimated contour in the DDSM. As the tumor sizes are not in constant ones; therefore, the size of ROI differs from each other. The example of ROI is shown in Figure 1.

Figure 1

(a) Training image (b) Region of interest from the training image

Morphological Enhancement

Traditional approaches for image enhancement are good at solving enhancement problems because they are from linear systems. Meanwhile, non-linear geometric approaches are needed in which mathematical morphology techniques are powerful nonlinear methodology.[14]

Some mathematical morphology enhancement techniques are applied in mass classification and segmentation mammography images. As the spicules are the important sign of malignant tumors, enhancing spinal axes of spicules is carried out by the use of morphological skeleton processing.[4] Another study is carried out with a set of morphological operation to remove the lesion background[5] and to enhance the disease patterns.

In this paper, structural morphology enhancement is studied with a new kind of primitive type. More information is presented in the section 3.

Feature Extraction and Selection

Mammography analysis needs the exact description of the ROI, and the texture features are common features in interpreting mammography images. These features are described by a set of statistical, structural, and spectral techniques. In this paper, the statistical and structural ones are used.

The statistical texture features from the ROI are gray level co-occurrence matrix (GLCM),[1516] gray level run length matrix (GLRLM),[17] and histogram statistics. In this kind of analysis, the features are calculated based on the statistical distribution of the pixel intensity in a given position toward the neighbors in a matrix of the image. The first feature, GLCM, is a table of frequency. In other words, it is the occurrence of pixel intensity value combination estimated in different directions as horizontal, 45, 90 and also 135. It should be noticed that texture is a pattern of pixel intensity in a special direction from a reference pixel. The run length matrix is counted as a number of pixels with the same intensity value in a distinct direction. The third feature describes the intensity distribution over the ROI. The estimated features are as follows: 22 features from the co-occurrence matrix, 11 features from run length matrix, and 6 statistical features. The names of the features are listed in Appendix A.

The structural feature is LBP (Eq. 1, 2). LBP is an operator to describe the spatial structure of an image in which the value of the central pixel (c) from a fixed size window is compared to its neighbors (in). Then the value of 0 or 1 is assigned to lower and upper value of the neighbors. These new assignments turn into a decimal number as the LBP feature.[18] Using a 3 × 3 window size produces a 256-bin histogram of the LBP labels.

There are some uniform patterns containing more information than others,[19] which are used here. Example of the local primitives including different types of curved edges, spots, and flat areas are shown in Figure 2.[20]

Figure 2

Example of texture primitive

Although the input image can be divided in a different window sizes, in this paper, the window is as the size of the ROI.

Proposed Approach

A kind of structural enhancement is proposed to increase breast mass classification performance in which the primitive of morphology operator has been modified. The overview of our method is shown in Figure 3. The new primitive is an optimized Gaussian kernel primitive as the kernel parameters are suited by the genetic algorithm. The proposed primitive is applied on both kinds of images, gray level (GL) and OD. Some texture features are extracted from these images, and finally the support vector machine (SVM) is used as a classifier. In the following scenario, optimizing the primitive elements and extracting features from the GL and OD images are described in details.

Figure 3

(a) The region of interest of the input image and (b) Illustration of the method

Morphology enhancement: Kernel based primitive estimation

Breast tissue has different structural elements. Although some enhancement techniques decrease these structural factors, they increase contrast enhancement. Therefore, a new method is proposed to raise the structural patterns diagnosis. For a better understanding, first morphological filtering theory on gray level (GL) primitive is described and then, the proposed primitive is discussed. As it was mentioned, the structural patterns are discoverable by LBP features; hence, variance of the LBP features is measured to justify the method, which comes at the end of this subsection.

The GL filters are as follows,[21] if f (x) is the image signal described on continues or discrete plane E = R2 or Z2 in which R =R̄∪{–∞,∞}, the most general translation-invariant morphological dilation of a GL image signal f (x) by another signal g is:

(ƒ⊕g)(x)≜Vƒ(x–y)+g(y)      (3)

The structural enhancement is designed by means of dilation operator and Gaussian kernel primitive. The primitive is a 7 × 7 window in which the elements are computed by 2-D Gaussian kernel.

In this way, all present elements in the primitive are not treated with the same costs. In other words, the primitive elements are weighting. The further elements take less weight, and those ones, which are spatially near the peak have more values. As dilation operator enlarges the maximum of the function,[21] and also breast masses appear lighter, which are the maximum gray levels in mammography images, applying Gaussian filters act like a hot spot; therefore, breast mass patterns get bolder.

An example of the primitive is shown in Table 1.

Table 1

Structuring element

Some examples of the enhanced images are displayed in Figure 4.

Figure 4

(a) The original image; (b) The enhanced image of A; (c) The original image and (d) The enhanced image

Variance is used as a measure of diversity, which means more variance causes more diversity on features. In the case of LBP features, more diversity means more patterns are discoverable. To justify the proposed method, one of the common enhancement methods is applied on the images. The common enhancement method adds the original image to the top-hat filtered image and then subtracts the bottom-hat filtered image. Then, the LBP features are extracted from both enhancement methods. Figure 5 shows more variance is achieved by applying the proposed method.

Figure 5

Variance calculated from local binary patterns features, dashed line for the proposed method and the strict line for the common enhancement method

Using OD images to remove GL background

The other proposed technique is to remove mass background with the exploit of OD mammography images. The proposed enhancement method, as mentioned in the previous section, is also applied on OD images-blackness of the processed film. Then a threshold method is assisted to remove the background. The eliminated pixels from the OD images are also removed from the gray level region of interest (GL ROI). As a result, some pixels that are more eligible to be mass are remained on GL ROI. In other words, background is eliminated from GL images. Then some texture based features as statistical, GLCM, and GLRLM are extracted from the GL ROI.

Classification

Among many classifiers, recently SVM is taken much attention in the classification of mammography masses.[92223] The SVM classifier goal is to map features into higher dimensional space while the feature vector is not separable in the original space. SVM classifier is used for both approaches with a linear kernel. For the test image, the two classifiers assign a label. The final decision is made by the maximum probability of the two classes in which the class with more probability assign the final label [Table 2].

Table 2

Ensemble classifier algorithm

RESULTS

In previous sections, the proposed approach was outlined in details. In this section, the classification performance for the two proposed methods on DDSM is evaluated and compared by using 5-fold cross validation and analyzing receiver operating characteristic (ROC) curve. In the following subsections, experimental results are shown as; in section A kernel parameter extraction procedure is described, in section B LBP feature classification on GL images are demonstrated, in section C some texture based features are extracted from OD and GL images, and section D combines the classifiers from two previous subsections and reports the results. However, first of all, the following terms are also used as the evaluation functions:

In this definition, malignant masses are mark as positive values and benign ones as negative. Therefore, false-positive (FP) and false-negative (FN) are two important factors. As FN answers may lead to the patients’ death, the FP answers lead to useless biopsy and bring pains for patients. The area under the ROC curve Az is also used as a factor for the classification performance.

Experimental 1: Kernel Parameter Estimation from GA

In order to select the appropriate Gaussian kernel primitive for the morphological enhancement, genetic algorithm is applied. The value of Mean for the 2-D Gaussian kernel is zero in both dimensions as the sigma parameter was learned by Genetic Algorithm. Only 30% of the input dataset is used for the training process. The fitness function is described by the LBP feature variance. The less variance causes more discrimination between the lesions.

Experimental 2: LBP Features on GL Images

The proposed primitive is applied on GL ROIs by dilation operator so the structural patterns are noticeable. Then LBP features are extracted from the GL images. In the case of comparison, the SVM classifies the extracted LBP features from dilated images with different kinds of primitive. The results are shown in Table 3, and its ROC is plotted in Figure 6.

Table 3

Result of SVM classifier on LBP features extracted from dilated images

Figure 6

Receiver operating characteristic plot for local binary patterns features

Experimental 3: GLCM, GLRLM and Statistical Texture Based Features on OD and GL Images

As it was mentioned, the second proposed method takes advantages of OD images. In this experiment, the proposed primitive is applied on OD images, and then Otsu's method is applied on the dilated OD images to set the threshold. This causes to remove the background tissue. Now the removed pixels from OD images are also removed from the GL images. This way, the pixels, which are more likely to be mass are remained. Then GLCM, GLRLM and statistical texture based features are extracted from the new images. The SVM classifier is also applied in this step. The table result is shown in Table 4, and its ROC is shown in Figure 7.

Table 4

Result of SVM classifier on texture features extracted from the eligible mass region

Figure 7

Receiver operating characteristic plot for texture features

Experimental 4: Ensemble the Classifiers

Now, the SVM classifiers from the two previous experiments are joined to get better results. The Final label for the test images is chosen by the maximum probability of the classifiers. Table 5 shows the SVM classifier results in joining the two classifiers from the OD images and their GL ones. Table 5 and Figure 8 show the performance results and the ROC respectively.

Table 5

Result of SVM classifier on LBP and texture features

Figure 8

Receiver operating characteristic plot for local binary patterns and texture features

DISCUSSION

All three experiments are evaluated by 5-fold cross validation, and other evaluation measurements, as mentioned, are classification accuracy, False positive rate, False negative rate, and Az. In the second experiment [Table 2], the LBP features are extracted from the original images and the dilated images with different primitives.

The results show that applying dilation operator on original mammography images with two common primitives is better than the pure original image. So dilating mammography images is a good idea. The proposed primitive works well, and classification accuracy rate is increased about 12%. The FNR is decreased about 11%, and also about 23% decrement is reported for the FPR.

In the third experiment [Table 3], the OD images are used to help GL images and remove some background pixels. The results show the proposed primitive works better than the common primitives, which are not good at dilating the OD images. The “original image” means the main GL image without any pixel removing. The present method results show removing some pixels affect about 2% improvement in the classification accuracy rate, and about 1% is decreased in both the FNR and FPR.

In the fourth experiment [Table 4], the two aforementioned features are joined to get a better result. As it is shown, the proposed approach is superior to the common primitives, and Az achieved to 92.31% in which about 2% increment is perceived. The best values for FNR and FPR is reached to 17.09% and 18.25% respectively. Although they are not noticeable, they are improved about 2% in value.

All in all, dilation operator is used to enhance the morphological patterns of the mammogram images. As it is shown in Table 2, some improvements by the common primitives are gained. Furthermore, the common primitives do not work well for OD images and causes negative effects on the overall classification. The proposed primitives are superior for all the 3 experiments results that show the effective process on OD and GL images.

Comparison of some mass classification methods are shown in Table 6. Many authors only cite the Az of their proposed method results without any attention to FNR and FPR while they are some important factors in the case of breast mass classification mammography. One of the recent researches by Amir Tahmasbi et al.[24] includes these factors. As it is shown in Table 6, the proposed method in this paper has a slight difference in FNR and FPR factors with this recent one, but performs better in Az[26].

Table 6

Comparison of some recent mass classification methods

CONCLUSION

Pre-processing is one of the most essential steps in many applications. In this paper, two new methods are applied; the proposed Gaussian kernel primitive and taking advantages from the OD images. Experimental results demonstrate their good performance on mass classification. The new primitives’ elements are trained with genetic algorithm; therefore, they are more powerful than the common primitives. Comparison of the proposed method results is somehow impossible with other works because of different sizes of datasets. As an example, the DDSM has 2620 cases, and only 516 of them are downloaded in this paper. Although we try to increase the classification accuracy rate and decrease FNR and FPR, they did not receive to an acceptable rate. To promote this work, we will add some shape descriptor features in proceeding paper.

BIOGRAPHIES

Azardokht Amirzadi received the B.Sc degree in computer engineering from Oloom Fonoon University, Babol, Iran in 2008. Currently she is doing MASc in Artificial Intelligence from Alzahra University, Tehran, Iran. Her research interests include machine learning and computer vision, medical image processing

E-mail: azardokht_amirzadi@yahoo.com

Reza Azmi received his BS degree in Electrical Engineering from Amirkabir university of technology, Tehran, Iran in 1990 and his MS and PhD degrees in Electrical Engineering from Tarbiat Modares university, Tehran, Iran in 1993 and 1999 respectively. Since 2001, he has joined Alzahra university, Tehran, Iran. He was an expert member of Image Processing and Multi-Media working groups in ITRC (From 2003 to 2004), Optical Character Recognition working group in supreme council of information and communication technology (From 2006 to 2007) and Security Information Technology and Systems working groups in ITRC (From 2006 to 2008). He was Project Manager and technical member of many industrial projects

E-mail: azmi@alzahra.ac.ir