A Novel Approach for Automatic detection of cancerous masses in mammogram MRI

A Novel Approach for Automatic detection of cancerous masses in mammogram MRI

(1)S.Pitchumani Angayarkanni, (2)Dr.Nadira

Banu Kamal,(3) Mrs.V.Thavavel Department of Computer Science, Lady Doak College,Madurai, Tamil Nadu, India Department of M.C.A, TBAK College

,Kilakarai,Ramnad,Tamil Nadu,India Department of M.C.A., Karunya

University,Coimbatore,TamilNadu,India

Breast cancer is one of the most common forms of cancer in women. In order to reduce the death rate, early detection of cancerous regions in mammogram images is needed. The existing system is not so accurate and also time consuming. The proposed system is mainly used for automatic segmentation of the mammogram images to classify them as benign, malignant or normal based on the decision tree ID3 algorithm. A hybrid method of data mining technique is used to predict the texture features which play a vital role in classification. Automatic classification is done through 3 stages ANN. The weights in ANN are adjusted using the rule derived from ID3 algorithm. The sensitivity, the specificity, positive prediction value and negative prediction value of the proposed algorithm accounts to 9.78%, 99.9%, 94% and 98.5% which rates very high when compared to the existing algorithms. This paper focuses on the comparative analysis of the existing methods and the proposed technique in terms of sensitivity, specificity, accuracy, time consumption and ROC.

Keywords: GLCM,Gabor filter, SOM,ANN and Data mining techniques.

1. Breast cancer has been determined to be the second leading cause of cancer death in women, and the most common type of cancer in women. The mammography is the

   best method of diagnosis by images that exists at the present time to detect minimum mammary injuries, fundamentally small carcinomas that are shown by micro calcifications or tumors smaller than 1cm. of diameter that are not palpated during medical examination. [Antonie et al, 2001]. Currently, joint efforts are being made in order to detect tissue anomalies in a timely fashion, given that there are no methods for breast cancer prevention. Early detection has proved an essential weapon in cancer detection, since it helps to prolong patients' lives. Physicians providing test results must have diagnostic training based on mammography, and must issue a certain number of reports annually. Double reading of reports increases sensitivity for detection of minimal lesions by about 7%, though at a high cost. The physician shall then interpret these reports and determine the steps to be taken for the proper diagnosis and treatment of the patient. For this reason, physicists, engineers, and physicians are in search of new tools to fight cancer, which would also allow physicians to obtain a second opinion [Gokhale et al, 2003, Simoff et al, 2002]. Different methods have been used to classify and/or detect anomalies in medical images, such as wavelets, fractal theory, statistical methods and most of them used features extracted using image processing techniques [1]. In addition, some other methods were presented in the literature based on fuzzy set theory, Markov models and neural networks. Most of the computer-aided methods proved to be powerful tools that could assist medical staff in hospitals and lead to better results in diagnosing a patient [Antonie et al, 2001]. Different studies on using data mining in the processing of medical images have rendered very good results using neural networks for classification and grouping. In recent years different computerized systems have been developed to support diagnostic

   work of radiologists in mammography[2].

   The proposed method includes the following phases i)Image Pre-processing and enhancement ii)Segmentation iii)Classification using ID3 Algorithm, iv) Predicting size and stages v) ANN and vi) Accuracy of algorithm prediction.

   PSNR	RMS	NSD	ENL	MES	Natur e of Filter
   87.65	2.97	4.55	89.89	8.83	Gabor

   Table 1: Signal to Noise ratio calculation

2. The main objective of pre-processing is to enhance the image and remove unwanted data. This is done by using gabor filter and histogram equalization. Gabor wavelet filters smooth the image by blocking detailed information. Mass detection aims to extract the edge of the tumor from surrounding normal tissues and background[1]. PSNR, RMS, MSE, NSD, ENL value calculated for each of 121 pairs of mammogram images clearly shows that gabor wavelet filter when applied to mammogram image leads to best Image Quality[4]. The orientation and scale can be changed in this program to extract texture information. Here 3 scales and 4 orientation was used[9].

   Input Image

   Output Image

   Histogram equalized

   Figure 1: Results after preprocessing and enhancement

   The results clearly show that the gabor filter with histogram equalization produces high PSNR value indicating that the image is highly

   enhanced.It also removes tape artifact, high intensity rectangular label and low intensity labels.

   2.1Comparison of Existing with proposed technique:

   Filters Used	Features & Limitations	PSNR value
   Median Filter	Used to smooth the non repulsive noise from two-dimensional signals without blurring edges and preserve image details. Suitable for enhancing mammogram images. Pectoral muscles are not detected	72.54
   LowPass Filter	Reduces noise and also blurs the edges	65.45
   Highpass filter	Enhances the details of the image	68.79
   Partial low and high pass filters	Best Quality image is achieved	84.90
   Spatial and frequency domain filters	Used for image enhancement alone	55.67
   Proposed Method	Features	PSNR
   Gabor Filter with histogram equalization	It acts as a local band-pass filter with optimal number of orientations and the scales define the number of filters that should	89.97
   	affect input images by
   	multiplying them with each other. The joint localization properties
   	of the image is enhanced by
   	histogram equalization in the
   	spatial and in the spatial
   	frequency domain. It is used for
   	detecting a first set of potential
   	microcalcifications and elongated
   	structures are identified It also detects clusters of
   	microcalcifications to extract
   	textural features of an image

3. Texture based segmentation is implemented because for a person affected by cancer the texture of the skin becomes smooth. This segmentation method segments the calcification pattern and the other suspicious regions in the mammograms The GLCM image is divided into 3×3 matrix and the texture features are calculated[2,3].

   Texture features are: Cluster prominence,Energy,Entropy,Homogenity,Differ ence variance, Difference Entropy, Information Measure, Normalized ,Correlation.

   Using GLCM (Gray Level Co-Occurrence Matrix) technique, the different combinations of brightness values that occur on the texture segmented image is found. Usually the GLCM matrix is found for small windows but in this project the GLCM matrix is found for the whole image. Then the GLCM Matrix is divided into small windows of size 3×3. Since the size of the Mammogram is larger, the size of the image is resized to 17×17 and the GLCM matrix gets segmented into 289 images. GLCM features: Cor-relation, Cluster Prominence, Energy, Entropy, Homogeneity, Difference Variance, Difference Entropy and Information Measure related to Cor-relation, and Normalized are calculated and stored in an Excel file.The texture values for 121 pairs of Mammogram MRI images are calculated and are stored in an excel sheet and it is analysed using SOM based Visualization technique .

   Pseudo code for performing Texture segmentation is:

   Step 1: Read Image

   Step 2: Create Texture Image

   Step 3:Create Rough Mask for the background Texture Step 4: Use Rough Mask to Segment the Foreground Texture

   Step 5: Display Segmentation Results

   Table 2: The Texture Parameter value for the left Mammogram Image Benign Case

   The unified distance matrix or U-matrix is a representation of the Self-Organizing Map that visualizes the distance between the network neurons or units. It contains the distance from each unit center to all of its neighbors. The neurons of the SOM network are represented here by hexagonal cells. The distance between the adjacent neurons is calculated and presented with different colorings. A dark coloring between the neurons corresponds to a large distance and thus represents a gap between the values in the input space. A light coloring between the neurons signifies that the vectors are close to each other in the input space. Light areas represent clusters and dark areas cluster separators. This representation can be used to visualize the structure of the input space and to get an impression of otherwise invisible structures in a multidimensional data space. The U-matrix representation (Figure 2) reveals the clustering structure of the dataset explored(Texture parameter) in this experiment. Texture parameters having similar characteristics are arranged close to each other and the distance between them represents the degree of similarity or dissimilarity.

   1. Visualization Output

   2. SOM TOOLBOX

      Figure 2: SOM based visualization for Benign case using SOM Toolbox

      The SOM toolbox produces secondary, strengthened, features which can then be used to segment or classify the image according to the texture energy. SOM toolbox in this research has helped to visualize the relationship between the features and also how the feature varies for different types of cases like Benign, Malignant and Normal.The output clearly indicates that Information Measure related to Correlation varies for the above specified cases during Mapping and it is also found that Energy and Entropy are oppositely correlated.

4. • A mathematical algorithm for building the decision tree.

   • Invented by J. Ross Quinlan in 1979.

   • Uses Information Theory invented by Shannon in 1948.

   • Builds the tree from top down, with no backtracking.

     Information Gain is used to select the most useful attribute for classification.

     Entropy

   • A formula to calculate the homogeneity of a sample.

   • A completely homogeneous sample has entropy of 0.

   • An equally divided sample has entropy of 1.

   • Entropy(s) = – p+log2 (p+) -p-log2 (p-) for a sample of negative and positive elements.

   • The formula for entropy is:

     Information Gain (IG):

   • The information Gain is based on the decrease in entropy after a dataset is split on an attribute[7].

   Information Gain(S,A)=Entropy(S)-H(S,A) Where H(S,A)=i(|Si|/|S|).H(Si)

   A takes on value 1 and H(Si) is the entropy of the system of subsets Si.

   The training data is a set S=s1,s2of already classified samples based on the texture features. Each sample Si=x1,x2 is a vector where x1,x2represents attributes or features of the sample. The training data is augmented with a vector C=C1,C2 and C3 where C1 represents benign, C2 represents malignant and C3 represents normal cases. At each node of the tree ID3 chooses one attribute of the data that most efficiently splits its set of samples into subsets enriched in one class or the other. The criterion is the normalized Information Gain that results from choosing an attribute for splitting the data. The attribute with highest information Gain is chosen to make decision.

   Pseudo code:

   Input: Set of texture feature attributes A1,A2—An. The class labels Ci.The number of classes Pi. Training set S of examples

   Output: Decision tree with set of rules.

   Node4	Homogenity
   Node5	Difference variance,
   Node6	Difference Entropy,
   Node7	Information Measure,
   Node8	Normalized
   Node9	correlation
   Class	Benign/Malignant/Normal

   Procedure:

   All the samples in the list belongs to three different classes

   Create a leaf node for the decision tree to choose the class.

   If none of the feature provides IG then ID3 creates a decision node higher up the tree using the expected value of the class.

   If instance previously unseen class is encountered, ID3 creates decision class higher up in the tree using expected value.

   Calculate IG for each attribute Choose attribute A with lowest entropy and Highest IG and test this attribute with root.

   For each possible value v of this attribute

   • Add a new branch below the root corresponding to A=v.

   • If v is empty make the new branch a leaf node labeled with most common value else

   • Let the new branch be the tree created by ID3

   Table 3: Nodes representing the 10 attributes

   Cross validation 10 fold

   === Detailed Accuracy By Class ===

   Class TP Rate FP Rate Precision Recall FMS ROC

   Benign 0.595 0.272	0.523	0.595	0.557	0.677
   Malig 0.121 0.104	0.368	0.121	0.182	0.542
   Normal 0.678 0.427	0.442	0.678	0.536	0.637
   Wgt. 0.465 0.268	0.445	0.465	0.425	0.618

   === Confusion Matrix === a b c <– classified as 172 30 87 | a = Benign

   94 35 160 | b = Malignant

   63 30 196 | c = Normal

   End

   The attributes used were the nine texture parameters with the class as benign, malignant and normal . Based on the rule derived by testing 121 pairs of various mammogram images the rules are applied in classifying the new cases without prior knowledge of whether they were benign, malignant or normal[8].

   An example of ID3 decision tree classification applied for a left benign case of mammogram MIAS dataset.

   Node1	Cluster Prominence
   Node2	Energy
   Node3	Entropy

   Error rate 0.5409

   Values prediction

   Value	Recall	1-Precision
   Benign	0.6090	0.4854
   Malignant	0.0000	1.0000
   Normal	0.7682	0.5771

   Confusion matrix

   Benign Malignant Normal Su Benign 176 0 113 2

   Infn Measure >= -0.0050 then Class =

   Normal (42.29 % of 525 examples)

   ID3 parameters:

   Size before split 200 Size after split 50 Max depth of leaves10

   Goodness of split threshold 0.0300

   Classifier performances

   This rule is then applied for classifying the new dataset of segmented mammogram images as benign , malignant and normal cases automatically. The results of the prediction were checked with the clinically proven classification results by the radiologist and were found that the rule derived provides 100%

   accuracy in classification.

   Maligna nt

   99 0 190 2

   S. No	No. of cases trained	No. of cases tested	Normal Trained/ tested	Benign Trained/ tested	Malignan t Trained/t ested
   1	40 Pairs	20 Pairs	3/6	25/4	22/10
   Table 4: No. of cases tested and trained

   Normal 67 0 222 2

   Sum 342 0 525 8

   Table 1: Predicting Cancer stages

   The size of the cancer is detected using the ellipsoid volume formula for the ROI using the formula

   The classification accuracy obtained from these fixed-order trees can be compared with those from trees of different feature orders, as well as with those from trees of different feature combinations. The decision tree with the optimal feature combination and order for this task can thus be identified. Examples of the training and test results obtained in this study will be discussed below. It should be noted that, to use a trained decision-tree classifier, one has to choose the specific tree structure with the set of decision thresholds corresponding to the desired sensitivity (TPF) and specificity (~- FPF).The structure and the thresholds will be fixed during testing or application. The decision rule is indicated below

   Infn Measure < -0.0450 then Class =

   Benign (64.89 % of 262 examples)

   Infn Measure >= -0.0450 then Class =

   Malignant (62.34 % of 316 examples).

   Statistical Method for stage prediction

   Cancer stage is based on four characteristics:

   the size of the cancer

   whether the cancer is invasive or non- invasive

   whether cancer is in the lymph nodes whether the cancer has spread to other

   parts of the body beyond the breast.

   /6 x L x W x H and 1/2 x L x W x H.

   Where L-length,W-Width and H-Height .

   Of the entire sample of women diagnosed with invasive breast cancer, 51% had stage I; 26% stage II; 11% stage III; and 4% stage IV disease. The equation predicting stage IV disease achieved sensitivity of 81%, specificity

   No. of Cases detected	Size	Lymph node	Stage
   51	<=2cm	Not Detected	I
   26	3 cm	Partially Detected	II
   11	6 cm	Fully Detected	III
   4	>6cm	Spread found other than lymph node	IV

   89%, positive predictive value (PPV) 24%, and negative predictive value (NPV) 99%, while the equation distinguishing stage I/II from stage III disease achieved sensitivity 83%, specificity 78%, PPV 98%, and NPV 31%. The equations most accurately identified early stage disease and ascertained a sample in which 98% of patients were stage I or II.

   suitable to deal with real problems which are nonlinear, non stationery and non Gaussian. The neural network classifier is used to generate a likelihood map of each mammogram using gabor feature as input to classifier.

   A neural network for solving classification problem typically has N input neurons and M output neurons. The kth output neuron (1<=K<=M) is trained to output one for pattern belonging to the kth class. A single output neuron suffices in the case of a problem that involves two categories of classification.ID3 based multilayered perceptron facilitates the classification of non linear problems .It is simpler to isolate the problem if the number of hidden neurons are high. The proposed method for the system is learning by trial and error which consists of adjusting weights of the connections according to the IG parameter value derived from ID3

   algorithm with ±Eq=1/2(rQ-o 2

   Table 5: Predicting Cancer stages

5. Multilayered feed forward neural network is used for training in this proposed method. Reason for choosing multilayer BPN is that it involves non parametric statistical properties. Unlike the classical statistical classification methods, such as bayes classifier

   ,no knowledge of the underlying probability distribution is needed by a neural network. It can learn the free parameters (weight and bias) through training by example. This makes it

   Q)

   Where rQ is Quadratic error among expected response and oQ is current response.

   The feature set extracted from 285 gabor feature set classified using the ID3 based Decision tree induction is used to classify benign, malignant and normal cases .These feature sets are given as input to the network for training. The desired output from the network is whether the classification is malignant, benign or normal tissue. Four neurons are used in hidden layer and the nine feature sets are used in input layer. During the training session of the network a pair of pattern is presented, the

   input pattern and target pattern (Malignant,Benign and Normal). At the output layer, the difference between the actual and target output yields an error signal. This error signal depends on the values of the weights of the neurons in each layer. This error is minimized and during these processes new values for the weights are obtained.

   Data partition results:

   Data partition results:

   575 records to Training set (68.05%)

   135 records to Validation set (15.98%)

   135 records to Test set (15.98%) Data anomalies:

   22 numeric outliers

   Network architecture: [9-4-3]

   Training algorithm: Online Back Propagation Number of iterations: 601

   Time passed: 00:00:06

   Training stop reason: All iterations done The best network was tracked and restored

   Figure 3: ANN Error Distribution

   Methods	Author and References	Computational Time
   Morphological Analysis	Wan Mimi Diyana, Julie Larcher, Rosli Besar	320
   Filtering Technique	Proposed Approach	350
   Fractal Dimension Analysis	Wan Mimi Diyana, Julie Larcher, Rosli Besar	720
   Complete HOS Test	Wan Mimi Diyana, Julie Larcher, Rosli Besar	920
   neuro-fuzzy segmentation	S. Murugavalli et. al	93 39
   Proposed Method	S.Pitchumani Angayarkanni ,Nadhira Banu Kamal	003

   Table 6: Comparatve analysis of Computational time.

ROC analysis is based on statistical decision theory, developed in the context of electronic signal detection, and has been applied extensively to diagnostic systems in clinical medicine. The ROC curve is a plot of the classifiers true positive detection rate versus its false positive rate. The false positive (FP) rate is the probability of incorrectly classifying a non-target object (e.g. normal tissue region) as a target object (e.g. tumor region). Similarly,

the true positive (TP) detection rate is the probability of correctly classifying a target object as being a target object. The TP and FP rates both are specified in the interval from 0.0 to 1.0, inclusive, in medical imaging, the TP rate is referred to as sensitivity, and (1.0 FP rate) is called specificity. Statistical classifiers have parameters that can be varied to alter the TP and FP rates. Using these parameters, an ROC curve can be generated which shows the TP/FP trade-off associated with the different values that the parameter(s) may assume. It would then be possible to trade a lower (higher) FP rate for a higher (lower) TP detection rate by choosing the appropriate value(s) for the parameter(s) in question.

Figure 5: ROC curve for Benign and Malignant case

Table 7: Accuracy of the proposed algorithm

The rule generated using decision tree induction method clearly shows that the time taken to classify benign malignant and normal cases in just 0.03 seconds and the accuracy is 99.9%.

The specificity =t-neg/neg, sensitivity =t-pos/pos and Precision=t-pos/t-pos+f-pos and accuracy=sensitivity pos/(pos+neg))+specificity(neg/(pos

+neg)). (Positive Prediction Value ) PPV: True positive / (true positives

+ false positives) PPV: 94% and (Negative Prediction value) NPV: True negative / (true negative + false negatives)=98.5%.From the above algorithm, the accuracy was found to be 99.9%. The proposed method yields a high level of accuracy in a minimum period of time that shows the efficiency of the algorithm.

So far the weights in ANN are not adjusted using decision tree rule evolved through data mining. This new concept has reduced the error rate and increased the efficiency.

References:

Table 8: Comparison of existing and proposed methods

Based Mammogram Image Retrieval. Journal of Applied Sciences, 9: 3531-3538

1. Randy P. Broussard et al, Gabor filtering for improved microcalcification detection in digital mammograms, British Journal of Radiology (2007) 80 2. Arnau Oliver, Jordi Freixenet, Joan Martí, Elsa Pérez,et al., A review of automatic mass detection and segmentation in mammographic images, Medical Image Analysis Vol. 14, Issue 2, Pages 87-110

1. Elodia B. Cole, Etta D. Pisano, Donglin Zeng, Keith Muller,et al. he Effects of Gray Scale Image Processing on Digital Mammography Interpretation Performance,

   Academic Radiology Vol. 12, Issue 5, Pages 585-595

2. Du-Yih Tsai,Yongbum Lee,Masaru Sekiya,Masaki Ohkubo, Medical image classication using geneticalgorithm based fuzzy-logic approach, Journal of Electronic Imaging 13(4), 780788 (October 2004)

3. T.J. Jose and P. Mythili, 2009. Neural Network and Genetic Algorithm Based Hybrid Model for Content

4. Maria-Luiza Antonie , Osmar R. Za¨ane

   , Alexandru Coman,Application of Data Mining Techniques for Medical Image Classification, International Journal of Computer Science & Information Technology (IJCSIT), Vol 3, No 1, Feb 2011

5. Aswini Kumar Mohanty, Saroj Kumar Lenka, Efficient Image Mining Technique for Classification of Mammograms to Detect

   Breast Cancer, Special Issue of IJCCT Vol. 2 Issue 2, 3, 4; 2010 for International Conference [ICCT-2010], 3rd- 5th December 2010

6. Aswini kumar mohanty, Sukanta kumar swain,Pratap kumar champati ,Saroj kumar lenka M, Image Mining for Mammogram Classification by Association Rule Using Statistical and GLCM features, IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 5, No 3, September 2011

7. Harald Fischer* and Ju¨ rgen Hennig, Neural Network-Based Analysis of MR Time Series Magnetic Resonance in Medicine 41:124131

8. L. Vibha , G M Harshavardhan ; K Pranaw ; P Deepa Shenoy ; K R Venugopal ; L M Patnaik , Classification of Mammograms Using Decision Trees, Database Engineering and Applications Symposium, 2006. IDEAS '06. 10th International,

pg263-266

Biography:

Iam an Assiatnt Professor in Computer Science at Lady Doak College.I received my B.Sc degree in Spl. Physics at Lady Doak College , M.C.A in Sri Meenakshi Govt. College for women and M.Phil in Madurai Kamaraj University.Currently Iam persuing my Ph.D at karunya University,Coimbatore.My areas of interest are Image Processing and Data Mining.

I am a Professor of the Department of Computer Science of TBAK College for Women, Kilakarai, Ramanathaputam District, Tamil Nadu,

India. I did my B Sc in Madras University, MSc in

Anna University, MPhil in Alagappa University and PhD in Manonmanium Sundaranar University. My area of Interests and research are Digital Image Processing and Software Engineering. I am at present supervising 5 research scholars for their PhD degree and have guided many MPhil scholars

Dr.Thavavel received M.C.A and M.hil Degrees from Madurai Kamaraj University, India. She is an Assistant Professor(SG) in M.C.A at Karunya University Coimbatore, India. Her research area of interest involves medical image reconstruction in an Object Oriented approach and application of Genetic algorithms to medical image analysis.