Reliable Computer-aided Diagnosis System using Region

Reliable Computer-aided Diagnosis System using Region based Segmentation of Mammographic Breast Cancer Images

BHAGWATI CHARAN PATEL1 Department of IT

Faculty of Engineering and Technology of Shri Shankaracharya Technical Campus Bhilai, INDIA

[email protected].

G.R.SINHA2 Department of ETC

Faculty of Engineering and Technology of Shri Shankaracharya Technical Campus Bhilai Bhilai, INDIA,

[email protected] ABSTRACT- Breast cancer is the most common type of cancer with women are affected in the world. Mammography is regarded as an effective tool for early detection and diagnosis of breast cancer. The most common breast abnormalities that may indicate breast cancer are masses and micro-calcifications or calcifications. The abnormalities present in breast images are characterized by an extensive range of features and may be simply missed or misinterpreted by radiologists while reading large amount of mammographic images during cancer screening process. Computer-aided diagnosis (CAD) systems have been developed by several researchers to assist radiologists provide an accurate diagnosis. We have attempted to improve the classification performance of CAD system in which shape and texture are used in analyzing region of interest (ROI) of mammographic images of breast. The method detects ROI by combining edge and region criteria and then feature extraction method helps extract few statistical parameters such as sensitivity and specificity to evaluate the performance of the proposed method. The sensitivity of the proposed method is 97.5% and specificity is 91.2% that produced an accuracy of 96.6%. Size of tumor is also computed and classification stage of breast cancer is identified. Keywords:- Breast cancer, Mammography image, CAD, ROI, Feature extraction, Sensitivity, Specificity 1 Introduction

Breast cancer cases have doubled in India in the last two decades. The number of women estimated to be dying of breast cancer every year has also been increasing. The cure rate of breast cancer if properly detected early is 97 percent but, unfortunately, less than 10 percent of all the 150000 new breast cancers diagnosed in India every year fall into this category. Despite the lower incidence of breast cancer in India than in the US, the number of women detected at an advanced stage of breast cancer is higher here. This is attributable to low awareness among Indian women on breast screening and self- examination [1]. Mmammography is the procedure of using low-dose X-rays to examine the human breast for identifying breast

cancer or other abnormalities present in breast images. Currently, for each patient that undergoes a mammogram, there is at least one X-ray image and one textual report written by a radiologist. The radiologist describes the features or structures in the report. If an abnormality or suspicious area is found, the patient may undergo a diagnostic mammogram or biopsy. Detection and diagnosis of breast cancer in its early stage enhances the chances for successful treatment and complete recovery of the patient. Finding an accurate and efficient breast region segmentation technique still remains a challenging problem in digital mammography [2]. Masses of breast cancer images have different density: low density, iso-densed, and high density; different

Mathematical Methods and Systems in Science and Engineering

ISBN: 978-1-61804-281-1 131

mailto:[email protected]

mailto:[email protected]

margins: circumscribed, micro-lobular, obscured, indistinct, speculated; and different shape: round, oval, lobular, irregular. Round and oval shaped masses with smooth and circumscribed margins usually indicate benign changes. Malignant mass has a speculated, rough and blurry boundary [3]. Calcifications are deposits of calcium in breast tissue. Malignant calcifications tend to be numerous, clustered, small, varying in size and shape, angular, irregularly shaped and branching in orientation [4]. Masses appear as dense regions of different sizes and properties [5]. Various spectrum of mammography masses are depicted in Fig 1. Depending on the morphology, the masses have dissimilar malignant property [6]. The ill-defined and speculated borders have higher probability of malignancy are shown in Fig. 2.

Fig 1. Types of mammographic masses.

(a)Circular shape(b)Lobular Shape(c)Speculated shape

Fig 2. Different shapes and borders of Masses..

2 Related Research Image segmentation is an important problem in image

analysis which is to partition a given image into disjoint regions. There are several image processing methods in current literatures for mass segmentation and feature extraction of breast cancer images. One of the commonly used methods is active contour model also called snakes method [7] whose basic idea is to start with a curve around object to be detected; and curve moves towards an “optimal” position and shape by minimizing its own energy. Snake or active contour is a curve in the image plane, which moves to minimize its internal and external energy functions. The internal energy function regularizes the shape of the contour. Based on the Mumford-Shah functional for segmentation [8,9], Chan et al. (2001) proposed a new level set model for active regions to detect objects whose boundaries are not necessarily defined by a gradient [10]. Contour-based information helps to overcome local minima and achieve a globally optimal solution irrespective of the initialization. Bresson et al. (2007) introduced a fast

method for global minimization of active contours after combining variational image segmentation and denoising method [11]. Roussain et al. (2007) used a convergence approximation for a multi-scale piecewise smooth model to overcome the limitations of global contour models while avoiding the high sensitivity of local approaches [12]. Cremers et al. (2008) and Leventon et al. (2008) utilized the shape to find a globally optimal solution, given the statistical shape is available[13,14]. After segmenting the suspected mass region, features of the segmented region are examined to verify whether the extracted region contains mass or not. Various features like intensity histogram, gray level co-occurrence matrix and intensity are used for breast cancer diagnosis. A key stage of mass detection and classification is feature analysis and extraction. Lie et al. (2001) implemented artificial intelligence (AI) techniques that include fractal dimension analysis, multi resolution Markov random field. The algorithm consists of three stages: fractal analysis, image segmentation and classification of suspicious regions. The fractal analysis is a pre-processing procedure. The classification stage uses six criteria to identify regions which are suspicious for tumor. The verification is made on the basis of sensitivity having obtained 97.3% and the number of false positive per image as 3.92 [15]. Shekhar et al. (2011) implemented a method for automatic breast cancer detection, classification, scoring and grading to help pathologists by providing second opinions. The classification of micro cancer object of breast tumor is based on feed forward back propagation Neural Network. The sensitivity, specificity and accuracy were found to be equal 93.10%, 95.70% and 92.60% respectively [16]. Pagonis et al. (2010) studied the effect of using multiple modalities on the accuracy achieved by a computer-aided diagnosis system, designed for the detection of breast cancer. A classification system was designed using the extracted features for every case. The classification accuracy increased at 82.95% using features from ultrasound that resulted increase in accuracy (95.12%) by using combined features from both x-ray and ultrasound [17]. Kowal et al. (2011) suggested a decision support system that allowed distinguish malignant from the benign breast tumors. The classification of the tumor is based on morphometric examination of cell nuclei. Features were extracted from segmented images obtained by hybrid segmentation method based on Gaussian mixture clustering and adaptive thresholding. Diagnostic accuracy obtained varies according to different classification methods and fluctuates up to 98% for quasi optimal subset of features [18]. We have proposed a method for easy identification of abnormal masses in mammographic images of breast


ISBN: 978-1-61804-281-1 132

having lower values of false positive (FP) and false negative (FN).

3 Pre Processing and CAD System Estimation of the volume any particular part of body

organ or objects within an organ is very important step in medical image analysis. For example, detection of tumor is clinically important in the analysis of medical image which could be achieved by estimating volume or the exact tumor area. The relative change in size, shape and the spatial relationships between anatomical structures obtained from intensity distributions help getting important information in clinical diagnosis. Therefore, radiologists are particularly interested to observe the size, shape and texture of the organs. For this purpose, organ and tissue morphometry is performed in every radiological imaging centre. A cancer cell has characteristics different from normal tissue cells in terms of cell outline, shape, structure of nucleus. Tumor, node and metastasis (TNM) is another stage of analysis for researchers to provide details about how the cancer looks and behaves [19]. Primary tumor (T) categories are:

• Tx: Tumor which cannot be assessed. • T0: No evidence of primary tumor. • Tis: No associated tumor mass. • T1 (includes T1a, T1b, and T1c): Tumor is 2 cm

(3/4 of an inch) or less across. • T2: Tumor is more than 2 cm but not more than

5 cm (2 inches) across. • T3: Tumor is more than 5 cm across. • T4: Tumor is of any size growing into the chest

wall or skin which includes inflammatory breast cancer.

If an area of suspicion is detected by a radiologist then a series of work-up procedure is recommended. The criteria for their decision are based on what they see and the location [20]. For example, radiologist finds a density then he/she has to decipher if the pathologic abnormality results as lesion or cancerous micro calcifications. If the answer is ‘Yes’, various procedure and modalities could be used such as core biopsy or ultrasound; and if it is ‘No’, additional views such as spot compression can be performed to see if the density is an architectural distortion, fibrosis or normal parenchyma. Fig. 3 shows the flow diagram of breast image examination. All the subsequent steps of the examination process are self explanatory.

Fig3. Breast image examination process.

High detection rate can be achieved by using suitable segmentation algorithm for detecting masses in mammograms; which includes three major steps: preprocessing, segmentation and post processing. The mammographic images need to be digitized prior the image processing. Fig. 4 shows various steps involved for the mass detection and feature extraction for breast cancer examination. The preprocessing step is used to improve the quality of the images subjected for subsequent stages: segmentation and feature extraction. Noise, if any and high frequency components in the mammography images are removed by using homomorphic filter. Edges are more important components in the segmentation of mammogram. The homomorphic filter removes the noise without disturbing the edges. Segmentation is important step in CAD system is the mass detection which aims to extract from the background or more regions of interest (ROI) which are likely to contain a mass. The objective of segmentation is to find out the entire suspicious mass region from mammogram.

In the feature extraction stage, the features are obtained from the characteristics of the region of interest. The best set of features is selected for eliminating false positives and for classifying lesion types. Fig. 4 highlights stages of mass detection and feature extraction of breast cancer image. Feature selection is made by selecting a smaller feature subset that leads to the largest value of some classifier performance function [19]. On the basis of selected features the false positive reduction and lesion classification are performed in the classification stage. Features of masses are distinct features from normal breast tissues.

Fine Needle aspiration

Breast image

Screening Mammogram

Diagnostic Imaging Mammogram, Ultrasound etc.

Normal Suspicious Equivocal

Palpable Mass

Suspicious Palpable Benign

Cyst Normal

Short term Follow up

Aspiration


ISBN: 978-1-61804-281-1 133

Fig.4. Mass detection and feature extraction.

Mammographic images are generally affected with various types of noise which can be removed without destroying the desired information; though it is a significant challenge. The filter is applied in frequency domain to the image which enhances the high frequency values and attenuates low frequency values. The filter function is defined as:

where r and c, number of rows and columns of the

image respectively; i and j, number of the rows and columns of each image pixel respectively.

Fig. 5 shows enhancement of a mammogram using homomorphic filter. The value of contrast was slightly decreased in comparison with that of original image (Table-1). The background noise level was decreased when compared with the original image. However the peak and average signal to noise ratio were slightly inferior to the ones of the original image. Mean (μ) and standard (σ) deviation are calculated as statistical measures:

(a) Original image (b) Filtered Image

Fig.5. Result of homomorphic filtering.

Table 1 EVALUATION OF HOMOMORPHIC FILTERING

Image μ Σ

PSNR(dB)

ASNR(dB)

Original image

5.47

3.58 45.78 32.5

Filtered image 4.970 3.233

4 Mass Segmentation Methods As already defined, a mass is space occupying lesion

seen in normal and abnormal mass. All suspicious regions are singled out for further analysis. The aim of the segmentation is to extract ROIs containing all masses and locate the suspicious mass candidates from the ROI. Segmentation of the suspicious regions on a mammographic image is considered high sensitivity and a large number of false positives are acceptable since they are expected to be removed in later stage of the algorithm [21]. It is difficult to segment image with complicated intensity distribution condition, or an inherent complexity in the object present in the image. Generally only one or two level set function are used to segment an image into region of two or more phase. If we use more level set function to segment complicated image into more regions, the dependency of the segmentation result on the initial condition is even more serious. In this work, we address both aspects of variation image segmentation and introduce an efficient and flexible numerical framework that enables us to compute a large class of advantageous descent directions.

Mammographic Image

Denoise the image

Perform Segmentation

Refine the edges

Segmented Mammographic Images

Feature Extraction

Evaluation of Segmentation

Feature Classification

Pre-processing Steps

Segmentation Steps

Post-processing Steps (CAD Evaluation)


ISBN: 978-1-61804-281-1 134

4.1 Previous Work Li et al (2008) proposed a method to overcome the

difficulties caused by intensity in homogeneities; a region-based active contour model was proposed that draws upon intensity information in local regions at a controllable scale. A data fitting energy is defined in terms of a region and two fitting functions that locally approximate the image intensities on the two sides of the region. This energy is then incorporated into a variation level set formulation with a level set regularization term, from which a curve evolution equation is derived for energy minimization [22]. Due to a kernel function in the data fitting term, intensity information in local regions is extracted to guide the motion of the region, which thereby enables our model to cope with intensity in homogeneity. Energy criterion

Where I(x) is the image intensity at pixel x, H is the Heaviside function, is a Gaussian kernel defined as:

With a scale parameter > 0. and are two functions centred at pixel x and computed at each iteration as:

and are two constants that have been set to 1 in the

interface. The two first integrals of correspond to data attached term, which are localized around each point x thanks to the Gaussian kernel . The third integral corresponds to the usual regularization term that smoothes the curve during its evolution. The last integral is a regularization term that forces the level-set to keep signed distance properties over the evolution process. Evolution equation is given as:

The parameters and have a fixed value of 1. Lankton et al (2008) proposed a natural framework that allows any region-based segmentation energy to be re-formulated in a local way. They considered local rather than global image statistics and evolve a region based on local information. Localized regions are capable of segmenting objects with heterogeneous feature profiles that would be difficult to capture correctly using a standard global method. They describe this framework and demonstrate the localization of three well-known energies in order to illustrate how our framework can be applied to any energy. They are then comparing each localized energy to its global counterpart to show the improvements that can be achieved [23]. Energy criterion:

(9)

where is the Dirac function, B is a ball of radius r centred at point x and defined as follow:

And

where is the Heaviside function, and are two functions updated at each iteration as follows:

The first integral of corresponds to a data attached term and the second is the usual regularization term that smoothes the region. Its evolution equation is given as:


ISBN: 978-1-61804-281-1 135

where and are the area of the local interior and local exterior regions respectively given by

4.2 Proposed Work

According to Li et al., we find the property of the localization introduced by and and . This algorithm is able to segment inhomogeneous objects and segments the whole image. is a signed distance function. This algorithm has three specific parameters that can be modified from the corresponding panel [22]: 1. The curvature term: it weights the influence of the

regularization of the evolving region (default value is set to 0.005).

2. The level-set regularization term : it weights the influence of the regularization of the shape of the level-set (default value is set to 1).

3. The variance of the Gaussian kernel (default value is set to 7).

Lankton et al suggested an algorithm that is a region-based method and its feature term is computed locally. This property allows the algorithm to segment non homogeneous objects. However this makes the method sensitive to initialization and is implemented as a signed distance function and is reinitialized at each iteration [23]. This algorithm has four specific parameters that can be modified from the corresponding panel: 1. The feature type: the user can choose feature type

(see equations (10), (14)). 2. The neighborhood type: the user can choose

between a circle or square neighborhood when computing the local statistics.

3. The curvature term : it weights the influence of the regularization term of the evolving region (default value is set to 0.2).

4. The radius term r: it allows to fix the radius size of the neighborhood (default value =9).

Key to our method is the notation of the probability-map F(x) derived from the original image I(x), so that F(x) has larger values at the features of interest. For

instance, if we are interested in the image edges, we can use a norm of the image gradient as a probability-map:

This notation is similar to the snakes' external energy, the higher probability areas of F(x) represent the desirable segmentation regions .We represent our active region , as a mean curve of the probability density function p(x):

p(x) can be seen as a continuous analog of Gaussian mixture model with equal isotropic covariance. We also add one more term to the p(x), a uniform distribution 1/N (where N is a size of the image domain), in order to account for noise and outliers. We rewrite p(x) as:

Parameter w represents the contribution of the uniform distribution (0 < w < 1). We search for the active region position, such that the distribution p(x) maximally approximates F(x), by minimizing Kullback-Leibler (KL) [24] divergence between the distributions:

which we call the fitness term. To regularize the shape of the region we keep the region regularization:

Which we call the regularization term. The goal is to find the function y(s) that minimizes the energy functional: E(y) = Efit(y) + Ereg(y) (22) The influence of each term is controlled by the weights α and β. We note that the probabilistic approach to active regions in Blake and Isard [25], where the mean curve is used to regularize the possible active region shape. In this experiments, we used α = β=1and also set the weight w of the uniform distribution term equal 0.1 to 0.9. The algorithm requires about 100 iterations to converge. We initialized the active region from the object. First, the active region converges to the mean of the probability-map. As we set to α, the active region enlarges to approximate the probability-map for a given scale α. As α decrease the active region achieves the optimum of its energy functional when it lies at the highest non-overlapping areas of the probability map and the curve produced by the active region is short and smooth. The active region does get attracted to the strait lines, but the shape regularization constraints forces to choose the smoothest and the shortest match. Here, we set β=5 to add some rigidity to the active contour. The active contour segments the boundary accurately, starting from


ISBN: 978-1-61804-281-1 136

an arbitrary location. We start from a large value of and gradually reduce it, tracking down the optimum. It does not guarantee to find the global optimum of the function, however empirically the optimum found is within a few percent from the global optimum. In the limit, when →0, it divergence converges to the conventional external energy term of active regions. For performing the comparison we used MSE and SNR parameter. Mean square error (MSE) is the cumulative squared error between original and resulted image. Signal-to-Noise-Ratio (SNR) measures are estimates of the quality of a reconstructed image compared with the original image. Experimental results show that our proposed method yields significantly improved visual quality as well as better SNR compared to the other techniques in the denoising. But in this approach our proposed method has better SNR and MSE to that of Li et al. but less than that of Lanktan et al. which are shown in Table-2. In region-based segmentation methods [10-11] the energy function depends on all image information, and thus operates on a global image scale. From this perspective, our method is global, because the fitness term depends on all image pixels and thus the method is not simply edge-based [25-26]. Fig. 6 depicts the various result for detecting the mass and edge. Once a ROI is detected, it is labeled as true positive ROI or false positive ROI. Draw a minimal rectangle fully containing the ROI, with its sides are parallel with respect to the ones of the image.

(a) Original image (b) according to Li

(c) According to Lankton d) Proposed method

(e) Define ROI (f) Edge & boundary

detection Fig. 6. Result of mass segmentation of breast cancer image.

Table 2

STATISTICAL MEASUREMENTS

S.N. Method MSE SNR

1 Li et al[22] 192.87 15.314

2 Lanktan et al[23] 138.65 22.759

3 Proposed method 162.34 20.976

5 Post Processing After the segmentation is performed on breast region,

the features can be obtained from it and the diagnosis rule can be designed to exactly detect the cancer nodules in the breasts. This diagnosis rules can eliminate the false detection of cancer nodules resulted in segmentation and provides better diagnosis. Correct detection of lesions, is essential to discover early breast cancer phases, increasing the treatment options and survival rate. In the feature extraction and selection step the features that characterize specific region are calculated and the ones that are important are selected for the classification of the mass as benign or malignant. The efficiency of a CAD system can be classified in four perspectives which are shown in Fig. 7 [27]:

• True Positive (TP), when the suspected abnormality is in fact malignant;

• True negative (TN), when there is no detection of abnormality in a healthy person;

• False positive (FP), when occurs detection of abnormality in a healthy person;

• False negative (FN), when there is no detection of a malignant lesion.


ISBN: 978-1-61804-281-1 137

Fig.7. Perspective of CAD system

The terminology which is used to determine the performance of a CAD System is defined as follows:

• Sensitivity: correctly classified percentage of ROI by radiologist. The sensitivity is the fraction of the true positive cases over the real positive cases:

High values of sensitivity imply minimal false negatives detection or higher true positive detection.

• Specificity: correctly classified percentage of ROI by non radiologist. The specificity of the test is the fraction of the true negatives case over the real negative cases:

High values of specificity imply minimal false positive detection.

• Accuracy: percentage of correctly classified pathological and non-pathological cases. Total accuracy shows the performance of the diagnostic system is determined based on the combination of sensitivity and specificity and is calculated by the following formula

Steps for calculating the various perspectives value of breast cancer in CAD system. Description: 1. Displays the observations, broken down by their actual and predicted groups. 2. Takes into account the chance that your predicted and actual groups may contain some mutually exclusive groups/ classes. 3. Assumes that union (predicted and actual groups) contains all possibilities for groups. 4. Returns the overall Accuracy and the following stats per group: True Positives, False Positives, True Negatives, False Negatives, Sensitivity, Specificity Input: 1. Predicted groups, the group for each observation, as predicted by your model. If you are using a logistic regression model, you need to translate the predicted

scores/ probabilities into groups, based on your own cut off value(s) and then feed those groups into this function. 2. Actual groups, the group for each observation, based on your actual data. 3. Arrange the group into Column and row wise. Where columns are assign different predicted groups, in ascending alphanumeric order, and rows are different actual groups, in ascending alphanumeric order. 4. Force both vectors to be column vectors and check equal length for each vector. 5. Set counter and matching the difference in column. If matched than set 0 otherwise 1. 6. Now get TN, TP, FP, and FN per class. Calculate Sensitivity, Specificity and accuracy. Output: 1. Display of the number of observations per group predicted as each group. 2. The overall Percent Correctly Classified in your data. Using these criteria, the results are usually defined in terms of Receiver Operating Characteristic (ROC) curve, which is the trade off between the true-positive rate and the false-positive rate inherent in selecting specific thresholds. To evaluate true-positive detection, sometimes is also required the localization of the tumor [28]. 6 Performance Evaluation We used a mammogram database developed by BSR Apolo, center for research and diagnosis of the database GRSDB, that are collected from local hospitals mentioning that biopsy has done on the patients, so we already know the results of benign or malignancy. There are 127 mammograms in the database taken from 89 different patients. The experiments are conducted on the proposed computer-aided diagnosis systems with the help of real time breast images. This experimentation data consists of 127 breast images. Those 127 breast images are passed to the proposed CAD system. The testing process is done for 56 cases. These 56 cases are fed to the proposed system and their output is recorded for calculation of the sensitivity and specificity. Using a small threshold is more likely to detect true lesions, but also to generate more false positive responses. Using a large threshold gives fewer false positive responses, but may miss more true lesions. Table-3 shows the results of quantitative analysis and from the results we can also prove the effectiveness of the proposed algorithm. Sensitivity, specificity and accuracy of prediction have been calculated according to the above formals for all of the testing data. The items of the tables include total accuracy (i.e., the percentage of correctly classified patterns), sensitivity (i.e., the probability that a case


ISBN: 978-1-61804-281-1 138

identified as malignant is indeed malignant), and specificity (i.e., the probability that a case identified as benign is indeed benign). The performance of the proposed method is evaluated using perfect test method which gives the sensitivity and specificity of the result with graphical representation. GRSDB-1 have highest sensitivity of the proposed method is 97.5% and specificity is 91.2% shown in Table-3. Hence the proposed method is highly desirable in order to assist the radiologist in the detection of malignant region and to improve the diagnostic accuracy. In Fig.8 binary image is obtained by thresholding the gray level than converted into pseudo colored and affected region are labeled with numbered from top to bottom and left to right . After finding the region than boundary is defined and brightest area are displayed.

Table 3

PERFORMANCE EVALUATION

Data Base Segmentat

ion method

TP FP FN TN S E

(%) S P (%)

AC (%)

GRSDB-1 Li et al. 1468 21 76 121 95.1 85.2 94.2

Lankton et al. 1478 19 65 153 95.8 89 95.2

Proposed method 1523 23 56 190 97.5 91.2 96.6

GRSDB-13 Li et al. 1358 29 83 135 94.3 82.3 93

Lankton et al. 1492 23 79 135 95 85.5 94

Proposed method 1392 37 69 156 95.3 80.8 93.6

Li et al. 1279 20 71 162 94.6 89 94.1

GRSDB-19

Lankton et al. 1479 27 73 152 95.2 85 94.2

Proposed method 1573 19 87 146 94.8 88.5 94.2

ROC curves illustrate the performance on a binary classification problem where classification is based on simply thresholding a set of scores at varying levels. Low thresholds give high sensitivity but low specificity, high thresholds give high specificity but low sensitivity; the ROC curve plots this trade-off over a range of thresholds. Using the results from Table-3 for the malignant cases, the receiver operating characteristic (ROC) curve of Fig. 9 has been extracted and reflects the system’s performance in terms of sensitivity and specificity (ROC analysis) for the malignant cases. The area under the ROC curve is a measure of the classification performance. A higher area indicates better classification performance because a larger value of TPR is achieved at each value of FPR. From Table-4, the general accuracy of the CAD system is “good” referring to their area under curve value of 96.6%. The accuracy of the test improves when the ROC curve moves closer

to the top left hand corner, that is, towards point (0, 1) of the graph [29]. Feature selection is thus useful for improving the result of accuracy in our experiment. Therefore, radiologists are particularly interested to observe the size, shape and texture of the organs and/or parts of the tissues. These routine assessments are commonly subjective and quantitative, and reports typically refer to lesions as large, small, and prominent. In Table-5 and Table-6 show the various numbers of segmented sub-area and their corresponding measurement respectively. The mammograms were digitized at 60μm per pixel (12-bit pixels) with a Lumisys 90 digitizer and averaged down to 0.1 mm per pixel. It can easily be seen that the size and diameter of the area which can be tumor are shown. Now we have to indentify the TNM classification.

(a)Binary image obtained by (b) Pseudo colored labeled.

thresholding

(c) Outline from boundary (d) Brightest area are displayed

Fig.8. Result of segmentation for define number of area

Table 4 AREA UNDER ROC CURVE (AUC) VALUES AND ACCURACY OF

THE DIAGNOSTIC TEST

AUC Value (%) Accuracy

90 – 100 Excellent 80 – 90 Good 70 – 80 Fair 60 – 70 Poor 50 – 60 Very

Poor


ISBN: 978-1-61804-281-1 139

Fig.9. ROC curve for GRSDB-39, depicting the performance

CAD system with sensitivity of 96.5% and specificity of 9.2%. Getting an accurate suspected region is a crucial issue because geometric features are extracted based on suspected regions, and these features are very important for further true lesion detection. The classification method proceeds as follows. As objects are isolated by the user, they may be given an area label. A count is maintained of the number of objects from each class that are represented in each Area. These counts are stored based on the scale factor; for instance, Table-7 shows the counts for the benign and malignant matched from Fig.1 and Fig. 2. We start counts by shape, size and the corresponding class. As Table-7 shows, size is a more important factor than shape for cell classification. An area number 6 and 9 gives the more prominent shape of speculated shape for tumor and their intensity is also higher than other segmented area. So they may be malignant but area number 5, 7,8,12 having the shape of circular or oval i.e. they represent the benign tumor. As on the area size is not more than 2 cm (3/4 of an inch) this tumor can be called as primary tumor of T1 category but cancerous cell is present. Our CAD system has been successful in indicating malignant micro calcifications seen mammographically and can improve the cancer detection rate during the screening. The use of a CAD system aids the radiologists as a second reviewer in evaluating mammograms and these systems are rapidly gaining acceptance in the breast imaging community. 7 Conclusions The aim of this paper is to improve the detection of suspected areas containing some type of lesion. Breast cancer is one of the major causes of death among women. Digital mammography screening programs can enable early detection and diagnose of the breast cancer which reduces the mortality and increases the chances of complete recovery. In the proposed work we have designed a new computer aided detection method to detect the mass region in the mammogram. Segmented image contains the suspected region which is given for feature extraction process. The extracted

features are classified in to normal and abnormal region .The obtained accuracy was 96.6% whereas the sensitivity and specificity were found to be 97.5% and 91.2% respectively. The proposed system gives fast and accurate classification of breast tumours. We discovered that there was a marked difference between the likelihood functions in malignant cases and the likelihood functions in benign cases. Quality measures including sensitivity, specificity, accuracy, PSNR and ASNR were calculated to determine the efficiency of the proposed method.

Table 5 SEGMENTATION INTO SMALL NUMBER OF AREA

Area #1 Area #2 Area #3 Area #4



Table 6 PARAMETERS FOR THE SEGMENTED AREA FOR GRSDB-1

(NUMBER OF PIXELS)

Area Mean Intensity Area Perimeter Centroid

Diameter

1 252.3 3.0 5.7 13.3 52 2.0 2 251 3.0 5,7 15.7 34.0 2.0 3 218.3 17.0 21.1 16.6 65.6 2.0 4 239.5 2.0 2.0 18.5 35.0 1.6 5 194.8 13.0 10.8 20.0 48.3 4.1 6 99.7 351.0 177.0 36.7 67.2 21.1 7 189.4 8.0 8.5 29.5 117.5 3.2 8 183.7 19.0 14.8 35.7 92.0 4.9 9 156.5 84.0 84.3 49.7 83.0 10.3

10 249.4 5.0 6.8 40.6 89.4 2.5 11 243.0 2.0 2.0 45.0 91.5 1.6 12 199.4 5.0 5.7 58.0 64.0 2.5


ISBN: 978-1-61804-281-1 140

Table 7

BENIGN/MALIGNANT COUNTS FOR MAMMOGRAPHIC IMAGES (SCALES AND SHAPES)

Image Benign

Malignant

GRSDB-1 4 2 GRSDB-13 6 1 GRSDB-19 4 3

ACKNOWLEDGMENT The authors extend their sincere thanks to Dr Dilip Soni, a senior radiologist for cancer research and diagnosis; for providing necessary support and suggestions throughout the research work.

REFERENCES 1. http://www.cancermission.com/2011/08/breast-

cancer-in-india.html 2. G.R.Sinha and B. C. Patel, Medical Image

Processing : concepts and applications, PHI Learning Private Limited,2014.

3. Cheng, H.D., Shi, X.J., Min, R., Hu, L.M., Cai, X.P., Du, H.N. Approaches for Automated Detection and Classification of Masses in Mammograms, Journal of Pattern Recognition, vol.39,issue.4, pp. 646–668,2006.

4. Rangayyan, R.M., Ayres, F.J., Desautels, J.E.L, A Review of Computer-Aided Diagnosis of Breast Cancer: Toward the Detection of Subtle Signs, Journal of the Franklin Institute,pp.312–348,2007.

5. Bruce, L. M., & Adhami, R. R. Classifying Mammographic Mass Shapes Using the Wavelet Transform Modulus-Maxima Method, IEEE Transaction on Medical Imaging, vol.1, pp. 1170-1177, 1999.

6. Arnau, O. Automatic mass segmentation in mammographic images, Thesis , ISBN :978846908019-1, Department of Electronics, Computer Science and Automatic Control, Girona, 2007.

7. D. Mumford, and J. Shah, Optimal approximation by piecewise smooth functions and associated variation problems, Journal of Pure Applied Mathamatics pp. 577-685,1989.

8. D. Mumford, and J. Shah, Boundary Detection by Minimizing Functionals, Proc. IEEE Conf. on

Computer Vision and Pattern Recognition, San Francisco,pp.176-185,1985.

9. J. Shah, Piecewise smooth approximations of functions, Calculus of Variations and Partial Differential Equations, vol.2, pp.315-328, 1994.

10. T. F. Chan, and L. A. Vese, Active contours without edges, IEEE Transaction on Image Processing., vol.10,pp. 266–277,2001.

11. X. Bresson, S. Esedoglu, P. Vandergheynst, J.-P. Thiran and S. Osher. Fast global minimization of the active contour/ snake model, Journal of Mathematical Imaging and Vision, vol. 28(2), pp.51-167,2005.

12. J. An, M. Rousson, and C. Xu. Convergence approximation to piecewise smooth medical image segmentation, In Medical Image Computing and Computer Assisted Intervention, vol. 4792, pp. 495-502, 2007.

13. D. Cremers, F. R. Schmidt, and F. Barthel. Shape priors in variational image segmentation: Convexity, lipschitz continuity and globally optimal solutions, In IEEE Conference on Computer Vision and Pattern Recognition, pp. 1-6, 2007.

14. M. E. Leventon, W. Eric, L. Grimson, and O. Faugeras. Statistical shape in_uence in geodesic active contours, In International Conference on Computer Vision, vo.1, pp. 317-323,2008.

15. Lei Zheng and Andrew K. Chan . An Artificial Intelligent Algorithm for Tumor Detection in Screening Mammogram, IEEE Transactions On Medical Imaging, vol.20(7),pp.45-52,2001.

16. Shekhar Singhand and Gupta, P. R. . Breast Cancer detection and Classification using Neural Network, International Journal Of Advanced Engineering Sciences And Technologies ,vol.6(1),pp. 004 – 009,2001.

17. Nikolaos Pagonis, Dionisis Cavouras, Kostas Sidiropoulos, George, Improving The Classification Accuracy Of Computer Aided Diagnosis Through Multimodality Breast Imaging, e-Journal of Science & Technology (e-JST), vol.2(5),pp.33-39,2010.

18. Marek Kowal, Paweł Filipczuk, Andrzej Obuchowicz, Józef Korbicz, Computer-Aided Diagnosis Of Breast Cancer Using Gaussian Mixture Cytological Image Segmentation, Journal Of Medical Informatics & Technologies vol.17,p.257-262,2011.


ISBN: 978-1-61804-281-1 141

http://www.cancermission.com/2011/08/breast-cancer-in-india.html

http://www.cancermission.com/2011/08/breast-cancer-in-india.html

19. http://www.breastcancer.org/symptoms/diagnosis/staging.jsp

20. Jain, A.K., Duin, R.P.W., Mao, J. Statistical Pattern Recognition: A Review, IEEE, Transactions on Pattern Analysis and Machine Intelligence vol.22(1),pp. 4–37,2000.

21. Bhagwati Charan Patel and G.R.Sinha, Energy and Region based Detection and Segmentation of Breast Cancer Mammographic Images, International Journal of Image, Graphics and Signal Processing,MECS Press, vol. 6,pp. 44-51,2012.

22. Chunming Li, Chiu-Yen Kao, John C. Gore, and Zhaohua Ding, Minimization of Region-Scalable Fitting Energy for Image Segmentation, IEEE Transactions on Image Processing, vo.17, pp.10-18,2010.

23. Shawn Lankton and Allen Tannenbaum, Localizing Region-Based Active Contours, IEEE Transactions on Image Processing, vol.17, pp.27-36,2008.

24. Kullback, S.; Leibler, R.A. On Information and Sufficiency , Annals of Mathematical Statistics, vol.22 (1),pp.79–86,1951.

25. Blake and Isard , Model Basedmulti-view Active Contours for Quality Inspection, Journal of Computer Vision and Graphics vol. 32,pp.565-574,2006.

26. R. Durbin, R. Szeliski, and A. Yuille. An analysis of the elastic net approach to the travelling salesman problem Journal of Neural Computation, vol.1,pp.348-358,1989.

27. Chenga, H.D.; Shana, J.; Jua, W.; Guoa, Y. and Zhangb,L. Automated breast cancer detection and classification using ultrasoundimages: A survey , Journal of Pattern Recognition, Elsevier Academic Press, vol.43,pp. 299–317,2010.

28. Sampat, M., Markey, M., & Bovik, A. Computer-Aided Detection and Diagnosis in Mammography. In A. Bovik, Handbook of Image and Video Processing, pp.1195-1217,2002.

29. B.C. Patel and G. R. Sinha , Mammography Feature Analysis and mass Detection in Breast Cancer Images, International conference on Electronic System , Signal Processing and ComputingTechnology(ICESC 2014), IEEE Conference, (2014) jan ,9-11,Nagpur,pp.474-478.


ISBN: 978-1-61804-281-1 142

http://www.breastcancer.org/symptoms/diagnosis/staging.jsp

http://www.breastcancer.org/symptoms/diagnosis/staging.jsp

http://link.springer.com/book/10.1007/1-4020-4179-9

Documents

Reliable Computer-aided Diagnosis System using Region